Agents Don’t Win. Rails Do.
Distribution, trust, and workflow not “agent builders” capture the value.
Written by Adrian Maharaj
Author’s note & disclaimer: I work at Google. The views here are my own and do not represent Google’s strategy, roadmap, or positions. This piece aggregates public information and operator experience.
The call: Agent marketplaces will be the hottest battleground in AI over the next 36 months and then the assembly layer commoditizes. Durable advantage accrues to whoever owns distribution, trust, and workflow integration. Everyone else becomes a feature.
B2B Two Sided Marketplaces The 90 Second Primer
A B2B marketplace matches supply (sellers/services) to demand (buyers) and compounds through cross side network effects: more quality supply attracts more buyers, which attracts more supply. The operating playbook is well known from the literature (Hagiu & Wright; a16z): seed a dense niche, solve the Cold Start, raise trust (identity, contracts, billing, SLAs), then scale into adjacencies. (Harvard Business School, ScienceDirect, Andreessen Horowitz)
Why this matters: mature marketplaces win because they don’t just list vendors they own procurement, billing, permissions, and compliance where work actually happens. That’s how the AWS Marketplace can credibly advertise 300,000+ active customers, 20,000+ products, and 3M+ subscriptions. And it’s why Salesforce AppExchange talks in install counts (13M+) and deep embeds, not just SKU pages. Distribution + trust = gravity. (Amazon Web Services, Inc., Salesforce AppExchange)
Independent analysts see this channel exploding: Canalys projects enterprise software sales via hyperscaler marketplaces to grow from $16B (2023) to $85B by 2028 a massive shift in how software is bought. (Canalys)
One more truth from app economies: power laws dominate. On mobile, the top 1% historically captured the overwhelming share of revenue; the long tail exists but rarely thrives without distribution. Expect similar dynamics for agents. (Sensor Tower)
What’s the Same and What’s Different About Agent Marketplaces
What stays the same
Liquidity math: You still need enough high quality supply (useful agents) meeting monetizable demand (outcomes). Cold starts, take rates, cross side effects still apply. (Harvard Business School, Andreessen Horowitz)
Trust decides: Identity/KYC, sandboxed permissions, secure billing, and observable performance separate rails from hobby shops. (Harvard Business School)
Distribution beats invention: Whoever already sits in the buyer’s workflow (clouds, suites, systems of record) can light up agent supply and monetize faster. See AWS and AppExchange examples above. (Amazon Web Services, Inc., Salesforce AppExchange)
What changes (and why it matters)
Agents are non‑deterministic. You’re listing behaviors, not static APIs. That forces marketplaces to ship evaluation harnesses, guardrails, and telemetry as first class features frameworks like LangGraph explicitly pitch “reliability and controllability” for production. (LangChain AI, LangChain)
Tool rights = bigger attack surface. With Anthropic’s Computer Use (desktop control), agents can click, type, and operate software. That’s powerful and raises governance stakes for any marketplace selling enterprise workflows. (Anthropic, Anthropic)
Regulatory pressure moves to the platform. The EU AI Act imposes risk based duties on providers and deployers, including documentation, logging, and oversight real obligations for marketplaces intermediating high risk workflows. (Digital Strategy, Artificial Intelligence Act)
Billing gets messy. Buyers don’t want to think in tokens + tools + compute. That favors marketplaces that abstract usage to an outcome price (“per dispute resolved,” “per ticket closed”) and plug into enterprise procurement like the cloud marketplaces already do. (Canalys)
Signals from the Giants—and What They Mean for SaaS Moats
Google (Gemini / Vertex)
Vertex AI Agent Engine/Builder targets enterprise deployment with managed runtime, IAM, and evaluation hooks. HIPAA‑aligned offerings on Vertex AI Search and RAG APIs show the direction: compliance baked into the rails. That’s a structural advantage when selling into regulated industries. Ya, call me biased, I am here for it. (General HIPAA guidance and Vertex compliance pages are public.) (Google Cloud)
OpenAI
The GPT Store made “agents as apps” visible to the mainstream. More importantly, the Assistants API is not priced separately—you pay for models and storage, not the control plane. Translation: the assembly layer is racing to utility pricing; your moat can’t be “we have an agent builder.” It must be distribution + trust + workflow lock‑in. (OpenAI)
Anthropic
Computer Use (public beta) lets agents operate a desktop great for back‑office ops, but it raises governance requirements for any marketplace allowing tool use outside narrow sandboxes. (Anthropic)
Microsoft/GitHub (the “agentic work” wedge)
Real‑world data: GitHub + Accenture reported developers code up to 55% faster with Copilot, and Microsoft’s Work Trend Index documents broad AI adoption inside large organizations. The lesson for agent marketplaces: when ROI is measurable, adoption moves. (The GitHub Blog, Microsoft)
Perplexity (distribution play)
Distribution matters: 780M queries in May 2025 shows how a query surface can steer demand to in house agents an aggregator risk for point solutions. (TechCrunch)
Counter‑signal (healthy skepticism)
Leading outlets note general‑purpose agents still miss the mark for broad consumer use, even as narrow use cases (coding, service ops) deliver. Costs are also non trivial for multi step, agentic tasks. This supports the view that vertical, controllable agents will win first. (The Verge, The Wall Street Journal)
What this means for SaaS moats
Feature moats die first. If your differentiation is “we added an AI feature,” incumbents with distribution will clone it.
Workflow + data + compliance harden. If your product is the workflow or controls sensitive data/perms, you’re safer longer (and can tax agent activity).
Trust becomes a SKU. Audit trails, policy engines, red‑teaming, signed execution—sell trust, don’t just message it.
Evidence the Future Is Already Showing Up
Cloud marketplaces are becoming the procurement rail: $16B (2023) → $85B by 2028 projected. AWS alone markets 300k+ active customers and 20k+ products proof that distribution + billing + trust compound. (Canalys, Amazon Web Services, Inc.)
Enterprise‑ready agent rails: Vertex exposes managed runtimes, evaluation, IAM, and HIPAA‑aligned services in the AI stack, signaling where agent platforms must go to win Fortune 1000 work. (Google Cloud)
App‑economy power law repeats: Most suppliers earn little; the head takes almost everything (cf. App Store revenue concentration). Expect a similar curve in agent marketplaces. (Sensor Tower)
Mainstream press framing: WSJ explicitly: “OpenAI Wants Businesses to Build Their Own AI Agents.” The enterprise agent wave isn’t hypothetical; it’s underway. (The Wall Street Journal)
Predictions (timestamp them)
2025–2026: Gold‑rush. Many agent marketplaces launch; most are thin “stores.” Revenue concentrates in a few measurable use cases (coding, CX ops, collections, sales ops). (See WSJ/FT coverage of agents entering everyday workflows.) (The Wall Street Journal, Financial Times)
2027–2028: Consolidation. 2–3 credible platforms emerge with distribution + compliance + procurement rails likely cloud marketplaces, productivity suites, and at least one vertical specialist in a regulated domain. (Tracks Canalys’ channel shift.) (Canalys)
By 2030: Rails commoditize. “Spin up an agent” becomes as cheap and boring as spinning up a container. Value accrues to: (1) owning the demand surface, (2) governing risky workflows, and (3) controlling proprietary context (data + workflows + partners).
Translation for a non‑technical CEO: agents are going to be everywhere; the winners won’t be the folks with the cutest bots, but the ones who control the storefront, the payment rail, and the rules.
Operator Playbook (for SaaS leaders)
Own the surface or the standard. If you can’t own the end‑user interface, publish schemas/ontologies and certify agents against them. Standards are soft moats that pull you into the center.
Be the programmable hub. Turn your product into a tool surface for agents: API first, outcome based billing, audit logs, tamper evident traces. Make agents travel through you then take rate.
Bundle trust. Ship policy engines, sandboxed tool scopes, signed execution logs, evaluation harnesses. Make risk a product and a pricing lever.
Verticalize now. Pick a regulated domain (health claims, trade finance, warranty adjudication). Build controllable agents with the compliance wrappers buyers demand.
Price outcomes, not tokens. Tokens are your COGS; customers buy results.
Measure like a marketplace: liquidity, match rate, time to value, repeat purchase, net GMV per cohort. (a16z’s network effects metrics are a solid dashboard.) (Andreessen Horowitz)
Selected sources & further reading
Marketplace mechanics & metrics: a16z on measuring network effects; Hagiu & Wright on multi sided platforms. (Andreessen Horowitz, Harvard Business School)
Cloud marketplace scale: AWS Marketplace (300k+ customers, 20k+ products, 3M+ subscriptions); Canalys forecast ($16B → $85B). (Amazon Web Services, Inc., Canalys)
App‑economy power law: Sensor Tower on top‑1% revenue concentration. (Sensor Tower)
OpenAI rails: GPT Store launch; Assistants API not priced separately. (OpenAI)
Enterprise‑grade rails: Vertex AI Agent Engine/Builder docs; Vertex AI Search compliance (HIPAA). (Google Cloud)
Agent capabilities & risk: Anthropic Computer Use (desktop control). (Anthropic)
Enterprise adoption signal: GitHub/Accenture on 55% faster coding; Microsoft Work Trend Index (AI at work). (The GitHub Blog, Microsoft)
Distribution shift: Perplexity at 780M queries in May ’25. (TechCrunch)
Regulation: EU AI Act, obligations for providers/deployers (risk‑based). (Digital Strategy, Artificial Intelligence Act).
