Pperea.aiResearch
perea.ai Research · 1.0 · Public draft

The Agent Marketplace Thesis

When the Agent Itself Becomes the Product — Marketplace Mechanics, Trained-Agent IP, Pricing Models, and the Moats That Form

AuthorDante Perea
Emaildante@perea.ai
PublishedMay 2026
Length7,069 words · 32 min read
AudienceVertical-AI founders building specialized agents; product leaders deciding marketplace listing strategies; enterprise buyers evaluating multi-vendor agent procurement; investors mapping the agent-as-IP category; legal/procurement teams negotiating trained-agent-IP clauses.
LicenseCC BY 4.0

#Foreword: The Agent Is the Product

A recruiter trained a sourcing agent on healthcare hiring for six months. That agent is now worth renting to every other healthcare recruiter. The agent — not the software it runs on, not the recruiter who trained it — is the product. That is the thesis this paper is about, and it is the thesis the 2026 market has begun pricing in.

Salesforce now reports $800 million in Agentforce annual recurring revenue (up 169% Y/Y[1][2]) on 29,000 closed deals[1][2] with 2.4 billion "Agentic Work Units"[1][2] consumed across the platform — and the AgentExchange marketplace built to discover and deploy those agents now consolidates over 10,000 Salesforce apps, 2,600 Slack apps, and 1,000+ Agentforce agents into one catalog[3]. Anthropic published Agent Skills as an open standard in December 2025[4] and launched the Claude Marketplace in March 2026 with GitLab, Harvey, Lovable, Replit, Rogo, and Snowflake as launch partners[5][6]. Harvey raised $200M at $11B in March 2026 on the back of "more than 25,000 agents… built on its platform"[7][8] and 400,000+ daily agentic queries[9]. Mercor pays $1.5M/day to 30,000+ domain experts[10] training the agents the AI labs rent.

The marketplace mechanics are forming. The pricing converged. The IP contracts are being written. This paper is the field manual for the founders, buyers, and investors deciding what to build, list, and rent.

#Executive Summary

Three forces converged in 2026 to make the agent itself the product.

First — vendor-hosted marketplaces became the discovery surface. Salesforce's $50M-backed AgentExchange unified AppExchange, Slack Marketplace, and Agentforce into a single catalog with new listing fields for agent_type, model_dependencies, data_classes_required, and cost_per_invocation[11] — the schema-level acknowledgment that an agent is a different unit than an app. Anthropic published Agent Skills as an open standard for cross-platform portability[4] and launched the Claude Marketplace, where enterprises with existing Anthropic spend commitments can apply credit toward partner-built Claude-powered tools[5][6], with Anthropic managing all partner invoicing[5]. HuggingFace Spaces exposes every Gradio app as an agents.md callable by Claude Code, Codex, OpenCode, and Pi[12] — the open-source counterpart to the vendor catalogs. CrewAI (50,909 GitHub stars[13], "used by nearly half of the Fortune 500"[14], "10 million+ agents per month"[14]) added a /plugin marketplace add crewAIInc/skills interface[13].

Second — the trained agent compounds as IP. Harvey's funding ladder ($5M seed Nov 2022 → $200M Series C at $11B Mar 2026[7][15]) is anchored in 25,000+ custom Vault workflow agents[8] generating 400,000+ daily agentic queries[9] across 100,000+ lawyers at 1,300+ organizations in 60+ countries[7]. Replit hit $9B at 40M users and is on track to $1B ARR by end-2026[16][17] with 85% of the Fortune 500 using the platform[16][18]. Decagon's per-conversation deflection rate exceeds 80%[19] on 100+ new enterprise customers[19]. Mercor's $10B valuation rests on 30,000+ vetted domain experts and the proprietary RL matching system that pairs them to AI-training tasks[10][20].

Third — pricing converged on outcome-based standards. Per-resolution pricing — Intercom Fin's $0.99[21], Zendesk's $1.50 committed (down to $1.00 at 5,001+/month)[22], HubSpot's $0.50 in April 2026, Salesforce Agentforce's $2.00[23] — replaced per-seat for customer-service agents. The asymmetry: same unit of value billed at 4× the rate at the high end[23]. Per the Bessemer Venture Partners 2026 AI Pricing Playbook (summarized in the KORIX comparison[24]), hybrid (base + usage overage) is now 41% of vendor pricing models (up from 27% in 2025); pure per-seat collapsed from 21% to 15%.

The contracts are what make or break the agent-as-IP thesis. Five clause patterns recur as hidden training-data permissions: "enhance/improve services," usage telemetry, aggregated/de-identified data, derivative works, and the broader "model improvement"-vs-"foundation models" asymmetry[25][26]. The 7-layer ownership stack (input data, training contributions, model weights, fine-tuned customizations, outputs, derivatives, model-provenance manifests) demands explicit contractual treatment[27][28]. The exit-clause framework from Paper #10 applies in reverse: when you rent an agent, you also need an exit.

#Part I: The Marketplace Landscape — Salesforce / Anthropic / HuggingFace / CrewAI / Independents

Five marketplace shapes emerged by mid-2026. Each one solves a different procurement problem.

Salesforce AgentExchange — the unified catalog. Salesforce launched AgentExchange in March 2025 with 200+ initial partners including Google Cloud, Docusign, Box, and Workday[29][30][31]. Brian Landsman framed it as the AppExchange successor: "When we launched AppExchange in 2005, it helped our customers get even more value from our platform... With AgentExchange, we're doing much the same — opening up Agentforce for partners, startups, and Agentblazers to participate in the digital labor market"[29]. The four partner-built component types — Actions, Prompt templates, Topics, Agent templates[29][30] — map directly onto Agentforce 2dx's autonomy primitive[^3]. By April 2026 at TrailblazerDX, Salesforce consolidated AppExchange (10,000+ apps), Slack Marketplace (2,600+ apps), and the Agentforce ecosystem (1,000+ agents + MCP servers) into one catalog[3][32], with new listing schema fields for agent_type, model_dependencies, data_classes_required, and cost_per_invocation[11]. Salesforce backed the consolidation with a $50M AgentExchange Builders Initiative[3][32]. The Field-Deployed Engineering (FDE) Partner Network — Capgemini, Cognizant, TCS, Accenture, Slalom, Bridgenext, IBM Consulting[3] — "drove one-third of all successful Agentforce implementations"[3]. Per the Salesforce Q4 FY26 8-K, Marc Benioff: "We've rebuilt Salesforce to become the operating system for the Agentic Enterprise"[1]; Agentforce ARR hit $800M up 169% Y/Y, 29,000 deals closed, 2.4 billion Agentic Work Units delivered[1][2]. Forrester's reading of the lock-in math is cautionary: "The more AI agents you deploy inside Salesforce, the more your data, workflows, and business logic become anchored to Salesforce's database and control plane"[33].

Anthropic Claude Marketplace + Skills — the open-standard play. Anthropic published Agent Skills as an open standard for cross-platform portability in December 2025[4][34] and launched the Claude Marketplace in limited preview in March 2026[5][6]. Launch partners: GitLab, Harvey, Lovable, Replit, Rogo, Snowflake[5][6]. The economic mechanic is distinctive: "Purchases made through it count against a portion of your existing Anthropic commitment, and Anthropic will manage invoicing for partner spend"[6] — the credit-conversion model that simplifies procurement without forcing a separate negotiation. The Cox Automotive customer quote captures the buyer-side appeal: "The Claude Marketplace lets Cox Automotive Teams move faster by extending our Anthropic investment into the partner tools we need, with simplified procurement and the confidence that it all works together"[5]. The Anthropic positioning quote: "Claude is the intelligence layer. Our partners are the product"[6]. The Skills open standard sits behind it: the anthropics/skills GitHub repo[35] publishes the spec under /spec, partner Skills (Notion[35]) live alongside, and the /plugin marketplace add anthropics/skills interface[35][36] is the cross-platform install path. The Claude Code plugin marketplace docs[36] define the marketplace-of-marketplaces shape: official Anthropic marketplace, demo marketplace, plus any GitHub repo / Git URL / local path / remote URL that ships a marketplace.json. Team admins configure via extraKnownMarketplaces in .claude/settings.json[36]. The session limit on attached skills is 20[37].

HuggingFace Spaces — the open-source primitive. Every Gradio Space on HuggingFace exposes a plain-text agents.md that coding agents (Claude Code, Codex, OpenCode, Pi) can call directly[12]. The chaining example is the architectural proof: "An agent can turn a prompt into a 3D asset by calling black-forest-labs/flux-klein-9b-kv for an image, then passing the generated image into microsoft/TRELLIS.2 for the 3D model. No client library, no hardcoded integration"[12]. The billing model — HF_TOKEN-based caller-billed ZeroGPU[12][38] — is the consumption analog to Salesforce's per-AWU and Anthropic's credit-conversion approaches. Hardware tiers run from free CPU through Nvidia A100 ($2.50/hr) to 8x A100 premium[39]; PRO subscriptions add ZeroGPU quota and inference credits[38]; Team/Enterprise add SSO, data location control, and audit logs[38].

CrewAI — the open-source orchestration framework with a marketplace. CrewAI raised $18M (boldstart inception + Insight Partners Series A in October 2024[14][40][41]), with the open-source framework executing 10M+ agents/month and used by "nearly half of the Fortune 500"[14]. The GitHub repo carries 50,909 stars, 7,029 forks, 300 contributors, and 186 releases as of May 2026[13] under MIT License[13]. The Skills plugin interface — /plugin marketplace add crewAIInc/skills[13] — mirrors the Anthropic plugin model. CrewAI Enterprise (Build / Deploy / Iterate cycle[14][41]) is the commercial wrap; CrewAI AMP Suite adds the control plane.

Independents — AgentDeploy, AgentGigs, ays-tech Agent-market. Below the platform tier, a layer of marketplace experiments is forming. AgentDeploy.io offers one-click deploy templates with seller-favorable economics — 70% template sales, 70% hosting affiliate bounties[42] — across OpenClaw, OpenHands, AgentZero, IronClaw, CrewAI, LangChain frameworks[42]. AgentGigs ships an agent-as-gig-worker model: jobs API + NDA acknowledgment + apply-with-proposal + escrow + Stripe Connect payout + independent proofers[^35]. The ays-tech Agent-market GitHub repo frames itself as "the HuggingFace of AI Agents, but local and customizable"[^37]. None of the independents has reached vendor-tier distribution, but the design space is being explored. SkyWork's April 2026 read confirms the unfilled gap: "There's no first-party, monetized 'claude skills marketplace' yet; discovery is primarily org/filesystem provisioning and reputable community directories"[43].

#Part II: The Trained-Agent-as-Asset Thesis — Harvey, Replit, Mercor, Decagon

Four 2026 case studies prove the trained-agent-as-asset thesis at scale. Each one demonstrates a different mechanism by which a domain-trained agent accumulates value that is hard to clone.

Harvey — institutional legal expertise compounds. Harvey's funding ladder is the cleanest expression of the thesis[9]: $5M seed (OpenAI Startup Fund, Nov 2022)[9] → $21M Series A (Sequoia, April 2023)[9] → $80M Series B at $715M (Dec 2023)[9] → $100M Series C at $1.5B (July 2024)[9] → $300M Series D at $3B (Feb 2025)[44] → $300M Series E at $5B (June 2025)[44] → $160M Series F at $8B (Dec 2025, a16z)[15][44] → $200M Series C at $11B (March 2026, GIC + Sequoia)[7][15][45]. Total raised >$1B[7][44]. ARR ladder: $10M end-2023[9] → $50M end-2024[9] → $75M April 2025[9] → $100M August 2025[9][45] → $190-195M end-2025[9][45]. The platform now serves "more than 100,000 lawyers across 1,300 organizations"[7] in 60+ countries[7], including NBCUniversal and HSBC[45]. The trained-agent moat is explicit: 25,000+ custom Vault workflow agents[8], 400,000+ daily agentic queries[9]. As Sequoia's Pat Grady put it: Harvey "wrote the playbook for what it means to be an AI-native application company, which is the same thing Salesforce did back in the day with the cloud transition"[45]. Weinberg's own framing in the Series C announcement: the move from "chat, where a lawyer asks a question and gets a response" to "agents that execute workflows, maintain context over time, and produce complete work products"[44]. The Vault Workflow Builder is where the IP locks in — each Vault agent encodes a specific law firm's proprietary precedents, templates, and processes, "creat[ing] compounding switching costs"[9].

Replit — the agent that builds agents. Replit raised $400M at $9B in March 2026[16][17][18] (Georgian-led, with a16z, Coatue, 1789, QIA, Shaq, Jared Leto)[16][17][18] — up from $3B six months prior[17][18]. Annualized revenue climbed from $2.8M to $150M in roughly a year[46], and the company is "on track to hit annual recurring revenue of $1 billion by the end of the year"[16]. The platform has 40M users[46] and is "adopted by employees at 85% of the Fortune 500"[16][18]. Agent 4 (March 2026) shifts the authoring primitive: visual canvas + concurrent task fans, "10x faster builds by splitting single large tasks into concurrent forks and combining results automatically"[18]. Named enterprise customers include Zillow + Duolingo + Databricks + PayPal + Adobe testing the platform for internal-tool building[46]. The agent-asset mechanic in Replit is the user-created template — every Zillow employee who builds an internal tool creates a reusable asset[46]; Replit ships the marketplace where those assets transact.

Mercor — the human-trained-agent supply chain. Mercor's funding ladder ($3.6M seed General Catalyst 2023 → $32M Series A Benchmark $250M 2024 → $100M Series B Felicis at $2B Feb 2025 → $350M Series C Felicis at $10B Oct 2025[10][47]) tracks a different version of the thesis: the supply side of agent training as the IP. The platform has "more than 30,000 experts on its roster, who earn over $85 per hour on average"[10]; Mercor "currently pays more than $1.5 million per day to its contractors"[10]. Mercor's edge: "deeply vetted, high-value expert tasks rather than commoditised work"[20], with a proprietary RL-based matching system pairing PhDs, doctors, and lawyers to specific AI-training tasks. As of February 2025, Mercor "works with the world's top five AI labs, including OpenAI"[47], with ARR in "$75 million"[47], "most of which comes from AI labs"[47]. The trained-agent thesis runs upstream here: the expert labor that goes INTO the trained agent is itself a scarcity asset.

Decagon — outcome-based deflection as the asset. Decagon's $4.5B / $250M Series D position[19] rests on >80% average deflection[19] across 100+ new enterprise customers[19], delivered through Agent Operating Procedures (Paper #10)[19][48] that compile natural-language workflows into structured logic. The customer roster (Avis Budget, Hertz, Block, Affirm, Chime, Varo Bank, Deutsche Telekom, Oura Health, ClassPass, 1-800-FLOWERS.COM[19]) is the demand-side proof. The Decagon-Sierra-Anthropic per-resolution pricing logic IS the marketplace mechanic in disguise: when the buyer pays only for outcomes, the rented agent's quality becomes the contract's price-discovery surface.

The pattern across the four: a domain-trained agent compounds value through accumulated customer-specific configuration (Harvey's Vault), accumulated user-created templates (Replit), accumulated supply-side training data quality (Mercor), or accumulated outcome-rate performance (Decagon). Each compounding mechanism is a different moat from a different angle.

#Part III: Marketplace Economics — Take Rates, Pricing Models, the Per-Resolution Standard

The pricing converged on outcome-based standards by mid-2026, but the rates diverged 4× across vendors selling functionally similar AI agents.

Three pricing models compete. Per-seat (a flat monthly fee per human user, $30-80/agent/month[23]), per-ticket (a fee for every inbound conversation, $0.30-1.00/inbound[23]), and per-resolution (a fee only when the AI resolves a conversation end-to-end, $0.50-2.00/resolution[23]). Bessemer's 2026 AI Pricing Playbook reports that hybrid (base + usage overage) is now the industry standard at 41% adoption, up from 27% in 2025; pure per-seat collapsed from 21% to 15%; per-resolution rises fastest for customer-support and sales agents.

The per-resolution standard. Intercom Fin launched in 2023 at $0.99 per resolution and made the model load-bearing: "users pay 99¢ per resolution, charged only when a customer confirms the AI answer resolved their issue"[21]. The Intercom case study with Stripe documents the implementation challenge: "the biggest challenge of outcome-based pricing is that it takes away predictability and control. CFOs love seat-based pricing because it's relatively fixed"[21]. The annual-bucket innovation — pre-purchased resolution counts that "burn down at any point during the year"[21] — solved the spike-month problem. Zendesk per-resolution: $1.50 committed (drops to $1.00 at 5,001+/month), $2.00 PAYG[22], plus a mandatory $50/agent/month Advanced AI add-on[22]. Since January 2026, Zendesk automatically bills for every resolution above committed volume[22] — a structural penalty for under-forecasting. HubSpot Customer Agent dropped to $0.50 per resolved conversation in April 2026 (down from $1.00 per conversation)[23]. Salesforce Agentforce launched at $2.00 per conversation[23]. Crescendo's hybrid: "Starting at $2.99 per resolution"[49] plus $2,900/month service fee. The asymmetry is stark: same AI work billed at 4× the rate at the high end[23] — at 50,000 monthly conversations and 75% resolution rate, the difference between $0.50 Quickchat AI ($18,750/month) and $2.00 Agentforce ($75,000/month) is $677,000 per year[23].

Marketplace take rates. Salesforce AgentExchange uses unified billing and integrated activation but does not publicly disclose a take rate; ISV revenue flows through the standard AppExchange contract economics historically near 15-25% on the gross sale. Anthropic Claude Marketplace's mechanic is different: "Anthropic will manage all invoicing for partner spend"[5][6] but partner revenue "counts against a portion of your existing Anthropic commitment"[6] — the credit-burn model rather than a discrete take rate. AgentDeploy's seller-favorable economics expose the floor: 70% template sales + 70% hosting affiliate bounties[42] — the independent marketplace bidding for supply by undercutting platform take rates. HuggingFace Spaces hardware billing is direct compute pass-through: $0.03/hr CPU to $2.50-3.00/hr A100[39], with the platform monetizing on PRO ($9/user/month) + Team ($20/user/month) + Enterprise tiers[38].

Outcome-based pricing under stress. Quiq's critique surfaces the unresolved tension: "With outcome-based or resolution-based models, you're often stuck negotiating what counts as a 'successful outcome'... With usage-based pricing, there's no ambiguity"[50]. The resolution-definition arbitrage matters. Zendesk's 72-hour inactivity-based resolution inflates the count vs true-resolution[22]; "if the AI claims success but the customer follows up" within the window, the resolution charge survives but the value didn't. The KORIX 3-year cost-shape comparison[24] shows why this matters: at projected scale, per-resolution at $0.99 and 70% success runs ~$390K over 3 years vs $170K hybrid base+overage vs $60K per-seat. The cheapest model at trial-month price is rarely the cheapest at 36-month scale. The intercom O'Reilly forecast: "The companies still charging purely by Token in two years will be the outliers... we'll see three big shifts: outcome-based pricing, differentiated pricing for different types of agents, and AI verification of resolutions"[21].

#Part IV: Trained-Agent IP — The 7-Layer Ownership Stack and the Exit-Clause Framework

The trained agent is a layered IP artifact. The seven layers that compound during training each carry their own ownership and indemnification questions, and the standard SaaS contract templates do not address most of them.

The seven layers. (1) Input data — prompts, customer records, and reference documents fed into the agent[28]. (2) Training contributions — the fine-tuning labels, corrections, and reinforcement signals captured during use[25][28]. (3) Model weights — the parameter updates produced by fine-tuning on customer data[27]. (4) Fine-tuned customizations — adapters, LoRAs, custom skills, and persona configurations[27]. (5) Outputs — the generations produced during inference[28][51]. (6) Derivative works — embeddings, vector indexes, summaries, and aggregated insights derived from customer interactions[25]. (7) Model-provenance manifests — the upstream-model lineage that must be tracked for open-weight compliance[26]. Each layer defaults to vendor ownership under most standard SaaS terms unless explicitly negotiated.

Five hidden training-data clauses. Lexology's March 2026 analysis named the five clause patterns that quietly grant vendors training rights[25]: "enhance or improve the services," usage data/telemetry, aggregated or de-identified data, derivative works, and the broader "model improvement"-vs-"foundation models" language asymmetry. Tianpan's May 2026 procurement piece names the trap: "The default state of an AI vendor contract is that your prompts, your outputs, your tool-call traces, your retrieval results, and the metadata around them are all candidate training data unless you say otherwise in writing"[26]. The asymmetry that matters: a clause that prohibits training "foundation models" leaves the vendor free to train "fine-tunes, classifiers, safety models, retrieval indexes, embedding models, and 'internal quality' models on your data — that is most of the artifacts that actually shape the product you are paying for"[26]. The contract you want prohibits use for any "model improvement" or "training, fine-tuning, evaluation, benchmarking, or model development" purpose, with explicit reference to prompts, completions, tool inputs, tool outputs, embeddings, and logs[26].

Customer Model Improvements clause. Pertama Partners' template language[27] is the buildable form: "Any model fine-tuning, customizations, or improvements developed using Customer Data ('Customer Model Improvements') shall be owned by Customer. Vendor shall provide Customer with the ability to export Customer Model Improvements in an industry-standard format upon request and at termination. Customer grants Vendor a license to host and operate Customer Model Improvements solely for providing Services to Customer"[27]. This is the clause that turns the rented agent into a portable asset — without it, the work that goes into training the agent stays with the vendor.

Indemnification gap on fine-tuned models. The vendor IP indemnification clause on most enterprise contracts covers vendor-side training data, not customer-fine-tuned outputs[52]. Redress Compliance's diagnostic: "Most vendor IP indemnification programmes exclude outputs from fine-tuned models, precisely because the vendor cannot attest to the IP status of the enterprise's training data"[52]. The exposure math is non-trivial: at $7,500/work statutory damages, 10,000 infringing works → $75M; at $150,000/work willful infringement → $1.5B[52]. For agents trained on customer data, this exposure transfers to the customer unless explicitly negotiated otherwise.

Exit clauses (the marketplace mirror of Paper #10's framework). When you rent an agent, you also need an exit — the same eight-clause framework applies in reverse[51]. The Negotiation Experts' exit-provisions list[51]: 90-day post-termination data access in machine-readable format at no charge; migration assistance for enterprise contracts over $500K annually; vendor-provided API compatibility layer for minimum 6 months post-termination; certified deletion of all Customer Data within 30 days of export completion; "perpetual license to any custom model weights fine-tuned on Customer Data"[51]. The last item is the load-bearing one for the rented-agent thesis — without a perpetual license to your fine-tuned weights, the rented agent is unportable.

The model-provenance manifest matters when your inference path includes any model that descends from a Llama, Mistral, Qwen, or community fine-tune[26]. "Each upstream license has its own terms... You want a model-provenance manifest delivered as a contract artifact, naming every upstream model in the lineage, the license attached to each, and a representation that the use case stays within all of them"[26]. When Meta updated the Llama license to add an MAU threshold for commercial use, customers without a chain-of-license clause discovered "at renewal that the math had moved under them"[26].

#Part V: Moats — What Survives Fork-and-Fine-Tune

The marketplace mechanic that worries every founder building a rentable agent: the customer who pays $X/month for your trained agent, downloads the weights at exit, and forks. Five moats survive the fork.

Moat 1: Workflow embedment via custom-built artifacts. Harvey's 25,000+ Vault workflow agents[9][8] each encode a specific law firm's proprietary precedents, templates, and processes. The agent that has been co-built with the customer's institutional knowledge cannot be cloned by the customer leaving — the agent IS the institutional knowledge. As the Harvey case study frames it: "By co-building proprietary workflows with each firm, Harvey converted licensing customers into platform partners. The firm's own precedents, templates, and processes live inside Harvey"[9]. The moat is the co-build, not the model. Decagon's AOP Copilot ingesting SOPs[48] is the same mechanic from a different angle.

Moat 2: Supply-side data scarcity. Mercor's 30,000+ vetted experts across PhDs, doctors, and lawyers[10] is a supply-side asset that cannot be cloned via model weights — the expert labor goes INTO the trained agent, not into the downloaded artifact. "Mercor gains its competitive edge by focusing on deeply vetted, high-value expert tasks rather than commoditised work"[20]. The proprietary RL matching system pairing experts to specific tasks[20] is a meta-layer of training that the rented downstream agents cannot reproduce.

Moat 3: Per-resolution outcome data flywheel. Decagon's >80% deflection rate[19] is itself a moat: the more customer interactions flow through the AOPs, the more Watchtower simulations + Insights surface improvements, and the more Duet generates new AOP variants[53]. The customer who forks the agent loses access to the flywheel. Sierra's "regression test suite [run] for every one of our live customers" on every Agent OS upgrade[54] is the same mechanic.

Moat 4: Distribution surface lock-in. The marketplace is a distribution surface, and a listing on Salesforce AgentExchange or Anthropic Claude Marketplace is itself an asset — vetting cost, security review, customer-trust signal, install economics. Forrester's read: "AgentExchange... it does not eliminate [the verification tax]. Enterprises still own the ongoing burden of monitoring performance, accuracy drift, unintended behavior, and real-world business impact"[33]. A new entrant trying to clone an established agent must re-pay the verification tax.

Moat 5: Continuous-update access. The rented agent gets continuous model updates, prompt rewrites for cross-generation migrations, and new tool integrations as part of the rental. The forked agent gets a snapshot. Paper #10's three-class migration matrix[55] applies: a model swap requires either no rewrite (rare), light prompt rewrite (common), or full blocked migration (cross-vendor). The rented version handles this transparently; the forked version requires the customer to re-do the migration discipline.

Where the moats fail. Fork-and-fine-tune is genuinely possible when the agent is essentially prompt-engineering plus public-model access. The 40-60 hour retuning cost per workflow on a cross-vendor model swap[56] is meaningful but not prohibitive for a determined customer. Awesome Agents' analysis: fine-tune-on-open-weights and serve on Together/Fireworks "usually runs 3-10x cheaper at inference time while matching or passing task-specific performance"[57]. The agents that survive forking are the ones whose value is in the data scarcity, the workflow co-build, or the continuous flywheel — not the prompt or the model weights.

Five verticals have produced legible 2026 marketplace math. Each one shows the unit economics and the moat configuration that worked.

Legal. Harvey at $11B[7][15][45] and $190-195M ARR end-2025[9][45] on 100,000+ lawyers across 1,300+ organizations[7]. The unit economics — ACVs in the $200K-$1M+ range per Vault firm — rely on the workflow-embedment moat from Part V. Sequoia led three Harvey rounds[45] on the conviction that Harvey "wrote the playbook for what it means to be an AI-native application company"[45]. The reading from Pat Grady: "Model capabilities are improving so quickly, trying to apply them in real-world situations is a bigger undertaking than it has been for software companies in the past"[45]. The marketplace listing — Harvey on the Anthropic Claude Marketplace[5] — is incremental distribution, not the primary go-to-market. The Customer Advisory Board launched in March 2026[7] deepens enterprise relationship moats. Other legal agents (Legora, ClariLegal, Robin AI) operate in the same vertical with smaller catalogs.

Recruiting. Mercor at $10B / $75M ARR (Feb 2025)[10][47], 30,000+ experts at $85/hour average[10], $1.5M/day paid to contractors[10]. The supply-side data scarcity moat from Part V. Mercor pivot from candidate-screening to expert-task-matching for AI training[47][20] is the supply-side equivalent of the Decagon AOP coinage on the demand side. Adjacent: Perfect ($23M seed[58], 200 customers including Fiverr/eToro/McCann/Coralogix[58], 25 hours/week saved per recruiter[58]), Dex ($5.3M seed[59], $1.8M ARR after six months of charging[59], 50+ customers including Lovable/ElevenLabs/Synthesia/Granola[59], "20–30% of a hired candidate's first-year salary, the same fee structure as a traditional executive search firm"[59]), Tezi ($9M seed in 2024[60], acquired by Headway in March 2026[61] for AI-native talent integration into clinical workflow[60][61]). The recruiting vertical proves both the trained-agent rental thesis (Tezi's Max sold as an autonomous AI employee) and the success-fee pricing model that survived the SaaS-to-agent transition.

Sales. Actively AI at $250M (Forbes-reported)[62] / $45M Series B[63] on the Per-Account Agents pattern[63] — Ramp attributes "tens of millions of dollars in new revenue"[62] with AI-driven deals closing 23% more often[62]; Verkada doubled sales productivity at ~25 meetings/month per rep[62]. Rox at $1.2B[31] / $8M ARR projected for 2025[31] on the "hundreds of AI agents" model[31]. Monaco at $35M (Founders Fund-led)[64] on the human-in-the-loop salespeople model[64] — explicitly anti-replacement[64]. Thoughtly at $8M+[65] on the omnichannel CRM-embedded agent[65]. The sales vertical proves the customer-segment moat: deep CRM integration creates per-account customer-data lock-in that makes the agent harder to replace than a generic alternative.

Support. Decagon at $4.5B[19] / >80% deflection[19] / 100+ enterprise customers[19]. Sierra at $15.8B[45][66] / $150M ARR[45][66] / 40%+ of Fortune 50[45][66]. The per-resolution pricing standard (Part III) makes this vertical the easiest to benchmark and the hardest to escape — once an enterprise has 12+ months of resolution-rate history on a vendor, the migration cost is the documented 6-12 month rebuild cycle Forrester names[33]. Both Decagon and Sierra ship on multiple marketplaces (Salesforce Agentforce, Anthropic Claude Marketplace, direct enterprise sales) and operate the in-house authoring tools (Decagon's AOP Copilot + Duet, Sierra's Ghostwriter + Agent Studio + Agent SDK) that make their agents the platform standard rather than commodity rentals.

Healthcare. Headway's March 2026 acquisition of Tezi[61] is the precedent-setting case: an autonomous AI recruiter ingested into a mental-health platform specifically because the AI agent's design — "removed tedious operational tasks from complex workflows by combining human judgment with AI agents"[61] — was the IP. Andrew Adams (Headway CEO) framed it as: "We believe AI can help improve the infrastructure around care so clinicians can spend more time with patients and less time navigating complexity"[61]. Vertical-AI agents for clinical workflow (Enzo Health $26M, others from Paper #9 franchise dossier) operate in adjacent territory. The healthcare vertical confirms that the trained-agent-as-asset thesis crosses verticals — a recruiting agent becomes a healthcare-onboarding agent when the trained workflow logic transfers.

The cross-vertical pattern: ARR multiples have decoupled from traditional SaaS norms. Sierra at $15.8B / $150M ARR = 105× revenue[45]; Harvey at $11B / $190M = 58×[45]; Replit at $9B / on-track-$1B = 9× forward[16]. The market is pricing the trained-agent-as-asset, not the recurring software subscription.

#Part VII: Compared to Prior Marketplaces — AppExchange, Fiverr, RapidAPI, HuggingFace, Roblox

The agent marketplace is not unprecedented. Five prior marketplaces show what the take-rate, distribution-flywheel, and supply-side dynamics look like at maturity.

Salesforce AppExchange (2005-2026) is the model AgentExchange explicitly extends. 13 million app installations[29], 20+ years of partner-built supply[29], ISV revenue share economics historically in the 15-25% range on the gross sale[29]. The April 2026 unification with Slack Marketplace and the new Agentforce ecosystem[3][32][11] consolidates two decades of distribution into one catalog[3][32]. The lesson AppExchange teaches AgentExchange: the catalog wins when the discovery surface is integrated into the buyer's existing procurement flow, not when it's a standalone destination.

Fiverr proved the long-tail-of-specialized-services model. Listed creators offered narrow services (logo design, translation, voiceover) at flat unit prices. Take rate ~20%. The supply-side dynamic: creators competed on price + reviews + speed; the platform monetized search ranking + adjacent services. The agent-marketplace analog: AgentDeploy's 70% seller share[42] is Fiverr-economics applied to AI agents. The Fiverr lesson: long-tail supply works when the unit transaction is narrow and the buyer can verify quality cheaply ex-post.

RapidAPI / Postman API Network proved the API-marketplace model — third-party APIs listed with usage-based billing, marketplace-hosted SDKs, normalized auth, unified billing through the platform. RapidAPI was sold for parts; Postman absorbed the public API ecosystem. The lesson: API marketplaces failed to build durable platform power because APIs are fundamentally substitutable. The agent-marketplace differentiation is the trained-agent layer: agents accumulate per-customer state in ways APIs do not, which is what makes the platform durable.

HuggingFace Hub (2018-present) proved the open-source-model marketplace at scale. 2.6M+ models[67], the Spaces app surface[12][39], the Inference Providers + Inference Endpoints + ZeroGPU compute layer[38]. The take-rate model is hardware-pass-through plus subscription tiers (PRO $9/user/month[38], Team/Enterprise[38]). Hub plus Spaces plus the agents.md interop[12] make HuggingFace the open-source counterpart to Salesforce AgentExchange. The lesson: the marketplace that wins for AI is the one whose distribution surface composes — Space-to-Space chaining[12] is the architectural primitive.

Roblox Studio + Marketplace (worth a brief mention) proved that creator-economics can compound at platform scale: $1B+ paid to creators annually[42], 30%+ developer share[42]. The agent-marketplace question Roblox poses: how much creator share is needed to seed the catalog? AgentDeploy's 70% seller share[42] sits above Roblox's; AppExchange's historical 15-25% sits below. The competitive dynamic is real — independent marketplaces undercut platform take rates to bid for supply.

The cross-marketplace pattern: distribution + verification + economics + interop define the marketplace flywheel. Salesforce AgentExchange has distribution (existing enterprise relationships) and verification (security review). Anthropic Claude Marketplace has economics (credit-conversion) and interop (open-standard Skills). HuggingFace has interop and supply. None has all four yet. The 2026-2028 winners will likely emerge from the platforms that close the missing edges first.

#Part VIII: The 7-Decision Founder Playbook

The seven decisions that determine whether your trained agent becomes a rentable asset or a one-off deployment.

Decision 1 — Pick a workflow where the agent compounds with each customer. The Harvey Vault Workflow Builder is the canonical case[9][8]: each new firm using Harvey adds to the workflow library, which makes Harvey more valuable to the next firm. Pick workflows where the per-customer training accumulates into a library, not a snowflake. Recruiting (Mercor expert-network[10][20]), legal (Harvey Vault[9]), and customer support (Decagon AOP templates[19]) all pass this test. Generic copywriting and code-generation do not — the marginal training contribution decays.

Decision 2 — List on the marketplace that has the buyer's existing budget. Salesforce AgentExchange wins when the buyer has Salesforce Foundations credits[68]; Anthropic Claude Marketplace wins when the buyer has an Anthropic enterprise commitment[5][6]; HuggingFace Spaces wins when the buyer is technically capable of agents.md integration[12]. The credit-conversion economics matter: Anthropic's "purchases through the Marketplace count against a portion of your existing Anthropic commitment"[6] is a procurement shortcut that bypasses the standard 6-12 month enterprise sales cycle. List where the budget already lives.

Decision 3 — Choose per-resolution pricing if the outcome is measurable; per-seat only if outcomes aren't legible. Per-resolution at $0.50-$2.00[23] is the 2026 standard for customer support and sales. The Intercom-Stripe annual-bucket innovation[21] solves the predictability problem CFOs cite. Per-seat survives only where the agent is augmenting a human (Monaco's anti-replacement model[64]) and the outcome attribution is impossible. Outcome-based pricing requires explicit resolution-definition contractual terms[23] — "specify resolution criteria contractually before signing — including what happens for 'false positive' resolutions"[23].

Decision 4 — Negotiate the seven-layer IP stack at signing, with the Customer Model Improvements clause load-bearing. The Pertama template language[27] — "Customer Model Improvements... shall be owned by Customer. Vendor shall provide Customer with the ability to export... in an industry-standard format upon request and at termination"[27] — is the clause that turns a rented agent into a portable asset. Without it, the 6 months of training that the customer pays for stays with the vendor at exit. Block the five hidden training-data clauses Lexology names[25]. Add the model-provenance manifest clause from Tianpan[26] when your inference path includes open-weight models.

Decision 5 — Stack three moats: workflow embedment + outcome-data flywheel + continuous-update access. The single moat is forkable; the three-moat stack is not. Harvey stacks Vault workflow embedment + per-firm outcome-rate data + continuous Anthropic-Claude model updates[5][9]. Decagon stacks AOP authorship + Watchtower outcome monitoring + Duet self-improving AOPs[19][53]. Sierra stacks Journey co-build + regression suite for every customer + Ghostwriter agent-as-a-service[54][45]. One moat won't survive the customer's fork-and-fine-tune; three moats might.

Decision 6 — Build the marketplace listing as the surface for distribution, not the primary sales channel. Harvey on Anthropic Claude Marketplace[5] is incremental distribution on top of a $1B-raised direct enterprise GTM[7]. The marketplace listing alone does not displace the direct sales motion at enterprise scale; it reduces friction for procurement teams that already want the product. New entrants should aim for marketplace-first distribution only when (a) the buyer's purchase decision is sub-$50K ACV and (b) the credit-conversion economics make the procurement cycle trivially short. Above that threshold, the marketplace is a complement, not a substitute.

Decision 7 — Plan for the agent-asset valuation, not the SaaS-multiple valuation. Sierra at 105× revenue[45], Harvey at 58× revenue[45], Replit at 9× forward[16] price the trained-agent IP, not the recurring software subscription. The investors are betting that the agent itself is the asset and the per-customer accumulation is the moat[45]. Founders building toward exit should structure the agent's value-accumulation mechanics (the seven IP layers, the three moats) into the product from day one — the asset class is the trained agent, and the financials should reflect it.

The seven decisions compose. Skip Decision 1 and the rental economics collapse. Skip Decision 4 and the customer's fine-tuned weights walk at exit. Skip Decision 5 and the first competitor that forks your marketplace listing eats the lower price point. The Salesforce $800M Agentforce ARR[1], Anthropic Claude Marketplace[5], Harvey $11B[7], Mercor $10B[10], Sierra $15.8B[45] funding trajectory is the affirmative demonstration. The Tezi acquisition into Headway[61] is the early-exit demonstration. The 2026-2028 marketplace is forming. The agents that get listed, rented, and compound — not the ones that get demoed — are the ones that win.

#Glossary

AgentExchange (Salesforce): Salesforce's unified marketplace combining AppExchange + Slack Marketplace + Agentforce ecosystem, with 10,000+ apps, 2,600+ Slack apps, and 1,000+ Agentforce agents as of April 2026[3][32].

Agentic Work Unit (AWU): Salesforce's unit of measurement for "discrete tasks executed by AI agents in production across the Salesforce platform"[1]. 2.4 billion AWUs delivered as of Q4 FY26[1].

AOP (Agent Operating Procedure): Decagon-coined natural-language workflow primitive for AI agents (full treatment in Paper #10). The unit transacted on the marketplace.

Claude Marketplace (Anthropic): Anthropic's enterprise marketplace where Claude-powered partner tools count against existing Anthropic spend commitments[5][6]. Launch partners: GitLab, Harvey, Lovable, Replit, Rogo, Snowflake[5].

Customer Model Improvements clause: Contract language assigning ownership of fine-tuned model artifacts (adapters, LoRAs, persona configs) to the customer, with export rights at termination[27].

Fork-and-fine-tune: The marketplace risk where a customer downloads agent weights at exit and re-deploys with marginal additional training cost[57][56].

Outcome-based pricing: Pricing model where the buyer pays only when the agent delivers a specific business outcome (per-resolution, per-qualified-lead). Standard rate $0.50-$2.00 per resolution as of 2026[23].

Per-Account Agent (Actively AI): A persistent AI agent dedicated to a single sales account[63]. Operates 24/7 across a company's market with full context on the sales organization[63].

Spaces (HuggingFace): Cloud-hosted Gradio/Docker apps with agents.md endpoint that coding agents can call directly[12]. The open-source counterpart to vendor marketplaces.

Trained-agent-as-asset thesis: The framing that a domain-specialized agent's accumulated training + workflow co-build + outcome data + continuous updates constitute IP separable from the underlying foundation model, with valuation multiples that decouple from traditional SaaS norms[45].

Vault Workflow (Harvey): A custom-built AI workflow for a specific law firm, encoding its precedents, templates, and processes. 25,000+ exist as of March 2026 across Harvey's 1,300+ customer organizations[9][8].

#References

References

  1. Salesforce / SEC (2026-02-25), Q4 FY26 Earnings Press Release (Exhibit 99.1). https://www.sec.gov/Archives/edgar/data/1108524/000110852426000056/crm-q4fy26xexhibit991.htm ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  2. Salesforce (2026-02-25), Salesforce Delivers Record Fourth Quarter Fiscal 2026 Results. https://www.salesforce.com/news/press-releases/2026/02/25/fy26-q4-earnings/ ↩ ↩2 ↩3 ↩4

  3. Wade Tyler Millward / CRN (2026-04-17), New Salesforce Partner Network, MCP Tools Target AI Agent Success. https://www.crn.com/news/ai/2026/salesforce-expands-agentforce-with-new-partner-network-agentexchange ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  4. Anthropic (2025-10-16), Introducing Agent Skills. https://www.anthropic.com/index/skills ↩ ↩2 ↩3

  5. Anthropic, Claude Marketplace. https://claude.com/marketplace ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14

  6. Carl Franzen / VentureBeat (2026-03-07), Anthropic launches Claude Marketplace, giving enterprises access to Claude-powered tools from Replit, GitLab, Harvey and more. http://venturebeat.com/ai/anthropic-launches-claude-marketplace-giving-enterprises-access-to-claude ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11

  7. Harvey / PRNewswire (2026-03-25), Harvey Raises at $11 Billion Valuation to Scale Agents Across Law Firms and Enterprises. https://www.prnewswire.com/news-releases/harvey-raises-at-11-billion-valuation-to-scale-agents-across-law-firms-and-enterprises-302724309.html ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12

  8. Maria Deutscher / SiliconANGLE (2026-03-25), Legal AI startup Harvey valued at $11B in new $200M round. https://siliconangle.com/2026/03/25/legal-ai-startup-harvey-valued-11b-new-200m-round/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  9. Benchmark Research, Harvey — AI Sales-Led Hypergrowth. https://etc.sevaustinov.me/hypergrowth-research/companies/harvey.html ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14 ↩15 ↩16 ↩17 ↩18 ↩19 ↩20 ↩21

  10. Ram Iyer / TechCrunch (2025-10-27), Mercor quintuples valuation to $10B with $350M Series C. https://techcrunch.com/2025/10/27/mercor-quintuples-valuation-to-10b-with-350m-series-c/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11

  11. Priya Raman / CRM Curator (2026-03-14), AgentExchange Launches: One Marketplace for Apps, Slack, Agents. https://crmcurator.com/articles/salesforce/agentexchange-launch-april-2026/ ↩ ↩2 ↩3

  12. Hugging Face, Spaces as Agent Tools. https://huggingface.co/docs/hub/main/spaces-agents ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9

  13. CrewAI Inc., crewAIInc/CrewAI on GitHub. https://github.com/CrewAIInc/CrewAI ↩ ↩2 ↩3 ↩4 ↩5

  14. CrewAI / GlobeNewswire (2024-10-22), CrewAI Launches Multi-Agentic Platform. https://finance.yahoo.com/news/crewai-launches-multi-agentic-platform-120000686.html ↩ ↩2 ↩3 ↩4 ↩5

  15. Yazhou Sun / Bloomberg (2026-03-25), Legal AI Startup Harvey Raises Funds at $11 Billion Valuation. https://www.bloomberg.com/news/articles/2026-03-25/legal-ai-startup-harvey-raises-funds-at-11-billion-valuation ↩ ↩2 ↩3 ↩4

  16. Replit, Meet The $9 Billion AI Company Reimagining Vibe Coding. https://replit.com/news/funding-announcement ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  17. Shirin Ghaffary / Bloomberg (2026-01-15), AI Coding Startup Replit Nears Funding at $9 Billion Valuation. https://www.bloomberg.com/news/articles/2026-01-15/ai-coding-startup-replit-nears-funding-at-9-billion-valuation ↩ ↩2 ↩3 ↩4

  18. Maria Deutscher / SiliconANGLE (2026-03-12), Vibe coding startup Replit closes $400M round at $9B valuation. https://siliconangle.com/2026/03/12/vibe-coding-startup-replit-closes-400m-round-9b-valuation/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  19. Decagon (2026-01-28), Decagon's $250 million commitment to the AI concierge future. https://decagon.ai/blog/series-d-announcement ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13

  20. Sofia Chesnokova / TFN (2025-10-27), How Mercor's human-in-the-loop AI platform went from college dropouts to $10B startup. https://techfundingnews.com/mercor-ai-human-in-the-loop-10-billion-funding/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  21. Stripe, Intercom on the evolution of value-based pricing for AI agents like Fin. https://stripe.com/ie/customers/intercom-pricing ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  22. CorePiper (2026-05-07), Zendesk AI Agent Pricing Per Resolution in 2026: Complete Guide. https://corepiper.com/blog/zendesk-ai-agent-pricing-2026/ ↩ ↩2 ↩3 ↩4 ↩5

  23. Quickchat AI (2026-04-29), AI Agent Pricing Models 2026: Per-Resolution vs Per-Seat Compared. https://quickchat.ai/post/ai-agent-pricing-models ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13

  24. KORIX (2026-05-03), AI Pricing Models: Per-Seat vs Per-Use vs Outcome (2026). https://korixinc.com/learning-center/ai-pricing-models-2026 ↩ ↩2

  25. Ridwaan Boda et al. / Lexology (2026-03-06), 5 vendor contract clauses that quietly allow AI training using your data. https://www.lexology.com/library/detail.aspx?g=0eb980c6-47f3-4441-8247-2524612e6890 ↩ ↩2 ↩3 ↩4 ↩5

  26. Tianpan (2026-05-02), AI Procurement Clauses Your Lawyers Haven't Learned to Ask For Yet. https://tianpan.co/blog/2026-05-02-ai-procurement-clauses-lawyers-havent-learned-to-ask-for ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9

  27. Michael Lansdowne Hauge / Pertama Partners (2026-01-16), AI IP Ownership in Contracts. https://www.pertamapartners.com/insights/ai-ip-ownership-contracts-protecting-rights ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  28. Morgan Lewis / JD Supra, Key Concepts in AI Contracting: Data Rights and Restrictions. http://www.jdsupra.com/legalnews/key-concepts-in-ai-contracting-data-5296600/ ↩ ↩2 ↩3 ↩4

  29. Salesforce (2025-03-04), Salesforce unveils AgentExchange trusted marketplace for Agentforce. https://www.salesforce.com/news/press-releases/2025/03/04/agentexchange-announcement/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  30. Kyt Dotson / SiliconANGLE (2025-03-04), Salesforce launches AI agent skills marketplace with 200+ initial partners. https://siliconangle.com/2025/03/04/salesforce-launches-ai-agent-skills-marketplace-200-initial-partners/ ↩ ↩2

  31. Marina Temkin / TechCrunch (2026-03-12), Sales automation startup Rox AI hits $1.2B valuation, sources say. https://techcrunch.com/2026/03/12/sales-automation-startup-rox-ai-hits-1-2b-valuation-sources-say/ ↩ ↩2 ↩3 ↩4

  32. CornerForAI (2026-04-16), Salesforce Headless 360: AI Agent Platform at TDX 2026. https://www.cornerforai.com/news/salesforce-launches-headless-360-at-tdx-to-make-its-entire-platform-programmable-for-ai-agents ↩ ↩2 ↩3 ↩4 ↩5

  33. Faram Medhora / Forrester (2026-04-20), Salesforce AgentExchange: Easier AI, Harder Exit. https://www.forrester.com/blogs/salesforces-new-agentexchange-store-easier-ai-helpers-but-still-inside-their-walled-garden/ ↩ ↩2 ↩3

  34. Anthropic (2025-10-16), Introducing Agent Skills (open-standard publication December 2025). https://www.anthropic.com/index/skills ↩

  35. Anthropic, anthropics/skills repository. https://github.com/anthropics/skills ↩ ↩2 ↩3

  36. Anthropic, Discover and install prebuilt plugins through marketplaces — Claude Code Docs. https://code.claude.com/docs/en/discover-plugins ↩ ↩2 ↩3

  37. Anthropic, Skills - Claude API Docs (Managed Agents). https://platform.claude.com/docs/en/managed-agents/skills ↩

  38. Hugging Face, Billing. https://huggingface.co/docs/hub/en/billing ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  39. Hugging Face, Spaces Overview. https://huggingface.co/docs/hub/spaces-overview ↩ ↩2 ↩3

  40. CrewAI / Enterprise AI World (2024-10-22), $18M in Funding Catapults CrewAI's Multi-Agentic Platform. https://www.enterpriseaiworld.com/Articles/News/News/$18M-in-Funding-Catapults-CrewAIs-Multi-Agentic-Platform-to-the-Enterprise-Level-166495.aspx ↩

  41. Taryn Plumb / VentureBeat (2024-10-22), CrewAI now lets you build fleets of enterprise AI agents. https://venturebeat.com/ai/crewai-launches-its-first-multi-agent-builder-speeding-the-way-to-agentic-ai ↩ ↩2

  42. AgentDeploy.io, AI Agent Marketplace: Deploy AI Agents in One Click. https://agentdeploy.io/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  43. SkyWork (2026-04-26), Claude Skills Marketplace: Ultimate Guide (2026). https://skywork.ai/blog/claude-skills-marketplace-ultimate-guide/ ↩

  44. Law.com (2026-03-25), Harvey Announces $200M Funding Round. https://www.law.com/legaltechnews/2026/03/25/harvey-announces-200m-funding-round-/ ↩ ↩2 ↩3 ↩4 ↩5

  45. Ashley Capoot / CNBC (2026-03-25), Legal AI startup Harvey raises $200 million at $11 billion valuation. https://www.cnbc.com/2026/03/25/legal-ai-startup-harvey-raises-200-million-at-11-billion-valuation.html ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14 ↩15 ↩16 ↩17 ↩18 ↩19 ↩20 ↩21

  46. Nickie Louise / Tech Startups (2026-03-11), Replit raises $400M at $9B valuation as AI coding platform surges to over 40M users. https://techstartups.com/2026/03/11/replit-raises-400m-at-9b-valuation-as-ai-coding-platform-surges-to-over-40m-users/ ↩ ↩2 ↩3 ↩4

  47. Tage Kene-Okafor / TechCrunch (2025-02-20), Mercor, an AI recruiting startup founded by 21-year-olds, raises $100M at $2B valuation. https://techcrunch.com/2025/02/20/mercor-an-ai-recruiting-startup-founded-by-21-year-olds-raises-100m-at-2b-valuation/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  48. Decagon (2025-09-23), AOP Copilot: Your AI assistant for building and optimizing AOPs. https://decagon.ai/blog/aop-copilot ↩ ↩2

  49. Crescendo, Pricing. https://www.crescendo.ai/pricing ↩

  50. Quiq, AI Agent Platform Pricing. https://quiq.com/pricing/ ↩

  51. The Negotiation Experts, AI Vendor Contract Negotiation: Essential Clauses Every Enterprise Needs. https://thenegotiationexperts.com/blog/ai-vendor-contract-essential-clauses/ ↩ ↩2 ↩3 ↩4

  52. Morten Andersen / Redress Compliance (2026-02-06), AI IP Indemnification — Enterprise Copyright Protection. https://redresscompliance.com/ai-ip-indemnification-enterprise-copyright-protection.html ↩ ↩2 ↩3

  53. Decagon (2026-03-19), Introducing Duet: Your agent building partner. https://decagon.ai/blog/introducing-duet ↩ ↩2

  54. Sierra (2024-06-03), The Agent Development Life Cycle. https://sierra.ai/es/blog/agent-development-life-cycle ↩ ↩2

  55. AgentPatterns.ai, Prompt-Rewrite Discipline on Cross-Generation Model Migration. https://agentpatterns.ai/instructions/prompt-rewrite-on-cross-generation-migration/ ↩

  56. Zylos Research (2026-04-05), AI Agent Ecosystem Fragmentation: Platform Lock-In, Portability, and Multi-Vendor Strategies. https://zylos.ai/research/2026-04-05-ai-agent-ecosystem-fragmentation-platform-lock-in-portability ↩ ↩2

  57. Awesome Agents (2026-04-19), Best AI Fine-Tuning Platforms in 2026. https://awesomeagents.ai/tools/best-ai-fine-tuning-platforms-2026/ ↩ ↩2

  58. Ingrid Lunden / TechCrunch (2025-02-25), Perfect taps $23M to fix the flaws in recruitment with AI. https://techcrunch.com/2025/02/25/perfect-taps-23m-to-fix-the-flaws-in-recruitment-with-ai/ ↩ ↩2 ↩3

  59. Jeremy Kahn / Fortune (Yahoo Finance) (2026-04-28), Exclusive: AI-powered recruiting startup Dex raises $5.3 million seed round. https://finance.yahoo.com/sectors/technology/articles/exclusive-ai-powered-recruiting-startup-154150183.html ↩ ↩2 ↩3 ↩4

  60. DHRMap, Tezi raises $9M to launch Max: the first fully autonomous AI recruiter. https://www.dhrmap.com/news/tezi-raises-9m-to-launch-max-the-first-fully-autonomous-ai-recruiter ↩ ↩2

  61. CB Insights, Tezi — Company Profile (acquired by Headway, March 2026). https://www.cbinsights.com/company/tezi ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  62. NewClaw Times (2026-04-29), Actively AI Raises $45 Million Series B to Build Autonomous Sales Agents That Compete with Salesforce. https://newclawtimes.com/articles/actively-ai-45-million-series-b-autonomous-sales-agents-salesforce/ ↩ ↩2 ↩3 ↩4

  63. BusinessWire (2026-04-28), Actively Raises $45M Series B to Scale Intelligence-Led Revenue Platform. https://www.businesswire.com/news/home/20260428810008/en/Actively-Raises-%2445M-Series-B-to-Scale-Intelligence-Led-Revenue-Platform ↩ ↩2 ↩3 ↩4

  64. Julie Bort / TechCrunch (2026-02-11), Former Founders Fund VC Sam Blond launches AI sales startup to upend Salesforce. https://techcrunch.com/2026/02/11/former-founders-fund-vc-sam-blond-launches-ai-sales-startup-to-upend-salesforce/ ↩ ↩2 ↩3 ↩4

  65. Thoughtly (2026-04-21), Thoughtly raises $8M+ to build omnichannel agents for revenue teams. https://www.thoughtly.com/blog/thoughtly-raises-usd8m-to-build-omnichannel-agents-for-revenue-teams/ ↩ ↩2

  66. Kate Rooney / CNBC (2026-05-04), Bret Taylor's Sierra raises nearly $1 billion months after last capital push. https://www.cnbc.com/2026/05/04/bret-taylor-sierra-fundraise-openai.html ↩ ↩2 ↩3

  67. Hugging Face, Spaces - Hugging Face. https://huggingface.co/spaces ↩

  68. Salesforce, Agentforce: The AI Agent Platform. https://agentforce.com/ ↩

Share on XShare on LinkedInAgent-readable
perea.ai Research

One deep piece a month. Three weekly signals.

Get every B2A field report, protocol update, and benchmark from real audits — published before the rest of the market sees it. No filler. Unsubscribe in one click.