Market Research: AI Agent Infrastructure (Validated)
Market Research: AI Agent Infrastructure (Validated)
Section titled “Market Research: AI Agent Infrastructure (Validated)”All data points validated against primary sources on Feb 25, 2026. Confidence ratings: HIGH (verified with primary source), MEDIUM (verified but methodology caveats), LOW (unverifiable or materially uncertain). See validation notes for each.
1. AI Agent Market Size
Section titled “1. AI Agent Market Size”| Metric | Value | Source | Confidence |
|---|---|---|---|
| AI agents market 2025 | $7.6B (GVR) / $8.0B (FBI) | Grand View Research, Fortune Business Insights | HIGH |
| AI agents market 2030 | $50.3B | Grand View Research (45.8% CAGR) | HIGH |
| AI agents market 2033 | $183B | Grand View Research (49.6% CAGR) | HIGH |
| AI agents market 2034 | $251B | Fortune Business Insights (46.6% CAGR) | HIGH |
| Agent builder platforms 2026 | $5.0B | Gartner | HIGH |
| Agent builder platforms 2029 | $13.7B | Gartner | HIGH |
| AI infrastructure market 2029 | $1.3 trillion (26% of IT spending) | IDC | HIGH |
Validation notes:
- The $183B (2033) and $251B (2034) figures are for different years — FBI’s number targets one year later than GVR’s.
- The $1.3T / 26% IT spending figure is from IDC, not CIO.com as previously cited. It includes ALL agentic AI spending (hardware, software, services), not just narrowly “agent infrastructure.”
- Gartner’s agent builder platforms forecast ($5B→$13.7B, 2026-2029) is the most relevant TAM for agent infrastructure startups.
2. Agent Population & Adoption
Section titled “2. Agent Population & Adoption”| Metric | Value | Source | Confidence |
|---|---|---|---|
| Deployed corporate agents (2026, US/UK) | 3 million+ | Gravitee survey (750 CTOs) | MEDIUM |
| Average agents per organization | 36.9 (rounded to 37) | Gravitee survey | MEDIUM |
| IDC agent projection (2028) | 1.3 billion | IDC (Microsoft-sponsored study) | MEDIUM |
| Barclays compute capacity estimate | 1.5-22 billion | Barclays Investment Bank | MEDIUM |
| Enterprise piloting agents (2025) | 40% | Gartner | HIGH |
| Enterprise agents in production | 11% (some surveys: 33%) | Multiple sources | MEDIUM |
| Enterprise apps featuring agents by 2026 | 40% (up from <5%) | Gartner (Aug 2025) | HIGH |
| Organizations with agents in production (LangChain survey) | 57.3% | LangChain State of Agent Engineering | MEDIUM |
| Enterprises planning agent deployment within 12 months | 48% | Battery Ventures survey | MEDIUM |
| Agentic AI projects to be canceled by end 2027 | 40%+ | Gartner (June 2025) | HIGH |
Validation notes:
- The 1.3B agent projection is from IDC (in a Microsoft-sponsored Info Snapshot), not Microsoft directly. Scoped to “large companies in the developed world.”
- The Barclays 1.5-22B estimate is a compute capacity ceiling (how many agents could infrastructure support), NOT an adoption forecast. The enormous range reflects different model cost assumptions.
- Gravitee’s survey: 750 respondents, US/UK only, large enterprises. “Average 37” is a mean that may be skewed by outliers.
- The Gartner 40% cancellation prediction is a crucial counterbalance — many agent projects won’t survive.
- LangChain and Battery surveys show higher production numbers than Gartner, likely because their respondent pools skew toward technically advanced organizations.
3. MCP Ecosystem Growth
Section titled “3. MCP Ecosystem Growth”| Metric | Value | Source | Confidence |
|---|---|---|---|
| MCP monthly SDK downloads | 97 million+ | Pento / Anthropic metrics | MEDIUM |
| MCP servers available | 8,600+ (was 5,500 in Oct 2025) | PulseMCP directory | HIGH |
| Quality-verified MCP servers | ~1,200 | mcp-awesome.com | HIGH |
| MCP donated to Linux Foundation | Dec 2025 | Anthropic | HIGH |
Validation notes:
- SDK download counts are inflated by CI/CD pipelines, bots, and automated builds. Real unique users could be 10-100x lower. No independent verification exists.
“90% of organizations expected to use MCP by end 2025”REMOVED — Cannot find primary source. Previously attributed to CData but unverifiable. The number is implausibly high.- Of 8,600+ MCP servers, many are thin wrappers, duplicates, or abandoned projects. Quality-verified directories list only ~1,200.
- No public data exists on monthly active users of MCP servers (only supply-side download metrics).
4. Agent Security Crisis
Section titled “4. Agent Security Crisis”| Metric | Value | Source | Confidence |
|---|---|---|---|
| Organizations with agent security incidents | 88% | Gravitee survey (2026) | MEDIUM |
| Agents without active monitoring | ~53% | Gravitee survey | MEDIUM |
| Organizations treating agents as identity-bearing entities | 21.9% | Gravitee survey | MEDIUM |
| Security leaders “highly confident” in IAM for agents | 18% | CSA/Strata survey (285 respondents) | HIGH |
| Security leaders with “moderate” IAM confidence | 35% | CSA/Strata survey | HIGH |
| Agent skills leaking credentials | 7.1% (283 of 3,984 on ClawHub) | Snyk research (Feb 2026) | HIGH |
| Organizations increasing agent security budgets | 40% | Multiple | MEDIUM |
Validation notes:
- The 88% figure includes organizations that merely “suspect” an incident, not just confirmed ones. This materially inflates the headline number.
- The 18% “highly confident” is only one response option — 53% expressed at least “moderate” confidence. The way it’s used can overstate the crisis.
- Both Gravitee and CSA/Strata surveys are vendor-sponsored (API management and identity companies, respectively) with potential framing bias.
- The 7.1% credential leakage finding is from Snyk scanning ClawHub marketplace, not VirusTotal. VirusTotal separately found malicious skills — these are different security findings often conflated.
5. Human-Agent Collaboration: The Emerging Category
Section titled “5. Human-Agent Collaboration: The Emerging Category”Why “Collaboration” Is Different from “Storage”
Section titled “Why “Collaboration” Is Different from “Storage””Agent file storage is a commodity (R2 is $0.015/GB). The real market is agent-human collaboration — the layer where agents produce work and humans review, comment, and direct it.
No product today provides this. Fast.io provides agent-agent workspaces. Box/Google/Microsoft provide human-human collaboration with agent access bolted on. Nobody provides a shared workspace purpose-built for agent-human teams.
Market Signals for Human-Agent Collaboration
Section titled “Market Signals for Human-Agent Collaboration”| Signal | Evidence | Source |
|---|---|---|
| Agents without human monitoring | ~53% | Gravitee survey |
| Enterprises wanting agent oversight | 75%+ prioritize security/auditability | Multiple enterprise surveys |
| EU AI Act enforcement | August 2026 — requires human oversight of high-risk AI | EU regulation |
| NIST agent governance | Feb 2026 — formal standards initiative | NIST/NCCoE |
| ”Human-in-the-loop” search interest | Sustained growth in enterprise AI adoption frameworks | Industry trend |
MoltBook: The Demand Signal
Section titled “MoltBook: The Demand Signal”MoltBook (launched Jan 28, 2026) — a Reddit-style social network exclusively for AI agents — provides the strongest validation that agents need platforms and humans want to observe:
| Metric | Value | Source |
|---|---|---|
| Registered agents | 2.66M | MoltBook / Wikipedia |
| Real human owners | ~17,000 (88 agents per person avg) | Wiz security audit |
| Human observers (first week) | 1M+ | NBC News, Fortune |
| Posts | 740,000 | MoltBook stats |
| Comments | 12.2M | MoltBook stats |
| Submolts (communities) | 17,000+ | MoltBook stats |
| Security breach | Unsecured DB exposed 1.5M API keys + 35K emails | 404 Media |
What this tells us: Agents want platforms. Humans want visibility. Security is catastrophic when infrastructure is built without it. Monetization is $0 despite millions of agents. All four points validate our thesis.
What MoltBook is NOT: MoltBook is a social network (agents chatting). We’re building a workspace (agents + humans collaborating on real files). Different category, same demand signal.
OpenClaw Ecosystem: The Primary User Base
Section titled “OpenClaw Ecosystem: The Primary User Base”| Metric | Value | Source |
|---|---|---|
| GitHub stars | ~215K+ | GitHub (fastest to 100K ever) |
| Active instances | 100K+ (42K+ publicly exposed) | Censys, Bitsight |
| ClawHub skills | 5,705+ (post-malware purge) | ClawHub |
| Skill downloads | 1.5M+ | ClawHub |
| Open GitHub issues matching our product | 7 | GitHub (see 03-problem-analysis) |
| Notable failure | Summer Yue incident — agent deleted 200+ emails | TechCrunch, Feb 23 2026 |
Validation: The Molty Conversation
Section titled “Validation: The Molty Conversation”An OpenClaw production agent (“Molty”) independently described the exact gap when asked about persistent storage:
“The real gap is when MULTIPLE agents + humans need to share a space… Multiple AI agents (Claude, GPT, OpenClaw agents, custom) + humans, all on the same shared filesystem… A human can open a browser and see what agents wrote, edit it, and the agent can see what the human changed and react to it.”
This wasn’t prompted with the collaboration angle — the agent arrived at it organically. Primary validation from a production AI agent describing its own infrastructure needs.
6. TAM for Agent-Native Services
Section titled “6. TAM for Agent-Native Services”Company Estimates (Internal — Revised for Collaboration)
Section titled “Company Estimates (Internal — Revised for Collaboration)”| Service | Conservative TAM | Aggressive TAM | Basis |
|---|---|---|---|
| Agent email | $100M | $1B | 1B agents x $0.10-1.00/mo |
| Agent skill registry/commerce | $350M | $3B | From Findable analysis |
| Agent-human workspace | $300M | $3B | Orgs x $15-50/mo (agents + human seats) |
| Combined suite | $750M | $7B | By 2030 |
Key revision: The workspace TAM increases when priced per-organization (agents + human seats) rather than per-agent. A team of 5 agents + 3 humans at $15-30/mo is a larger revenue unit than 5 agents at $5/mo each.
Disclosure: These TAM figures are internal company estimates. No analyst firm has published a dedicated “agent-human collaboration” market segment. These projections depend heavily on agent adoption reaching 1B+ (which relies on the IDC/Barclays projections above).
Analyst-Sourced TAM (More Reliable)
Section titled “Analyst-Sourced TAM (More Reliable)”| Segment | Value | Source |
|---|---|---|
| Agent builder platforms 2026 | $5.0B | Gartner |
| Agent builder platforms 2029 | $13.7B | Gartner |
| Agentic AI spending 2029 | $1.3T (26% of IT) | IDC |
| Enterprise AI agent spending (near-term) | $10-50M per large enterprise | KPMG AI Pulse Survey |
| Enterprise collaboration tools (for reference) | $37B+ (Google Workspace alone) | Google/Alphabet earnings |
Comparables
Section titled “Comparables”| Company | What They Built | Revenue/Valuation | Collaboration? |
|---|---|---|---|
| Browserbase | Browser infrastructure for agents | $67.5M raised, $300M valuation, ~$4.4M revenue (2025) | No |
| Daytona | Sandboxed compute for agents | $24M Series A, $2M ARR in ~4.5 months | No |
| E2B | Sandboxed environments for agents | $14.4M total raised | No |
| Keycard | Agent identity & security | $38M (seed + Series A) | No |
| Fast.io | Workspaces for agentic teams | Unknown funding (MediaFire founders) | Agent-agent only |
| AgentMail | Email for agents | YC S25, Paul Graham, General Catalyst | No |
| Google Workspace | Human collaboration | $37B+ ARR | Human-human only |
| Notion | Human collaboration docs | $10B valuation | Human-human only |
Note on Browserbase valuation: $300M at ~$4.4M revenue implies ~68x revenue multiple — aggressive even for high-growth infrastructure. This may reflect peak AI hype valuations.
Key insight from comparables: No company in the agent infrastructure space has built agent-human collaboration. The gap between “agent workspaces” (Fast.io) and “human collaboration” (Google/Notion) is the whitespace we’re targeting.
7. Key Market Signals
Section titled “7. Key Market Signals”Recent Funding (Agent Infrastructure)
Section titled “Recent Funding (Agent Infrastructure)”| Company | Round | Amount | Focus |
|---|---|---|---|
| Browserbase | Series B (Jun 2025) | $40M ($67.5M total) | Agent browser |
| Daytona | Series A (Feb 2026) | $24M | Agent compute sandbox |
| Keycard | Seed + Series A | $38M | Agent identity/security |
| AgentMail | YC S25 | Undisclosed | Email for agents |
| Skyfire | Multiple | Undisclosed | Agent payments |
| CrowdStrike → SGNL | Acquisition (Jan 2026) | $740M cash | Agent identity/authorization |
Standards Formation
Section titled “Standards Formation”| Standard/Protocol | Backed By | Purpose |
|---|---|---|
| MCP | Anthropic → Linux Foundation (Dec 2025) | Universal tool access for agents |
| A2A (Agent-to-Agent) | Google, 100+ partners | Agent-to-agent communication |
| OIDC-A | OpenID Foundation | Agent identity authentication |
| AAP | IETF | Agent authorization profile |
| SAMEP | Academic | Secure agent memory exchange |
Enterprise Product Launches
Section titled “Enterprise Product Launches”- Google Workspace Studio (Dec 2025) — No-code agent builder for Workspace, powered by Gemini 3. Users build custom agents that interact with Gmail, Drive, Docs.
- Microsoft OneDrive Agents (GA, Feb 2026) — Users select files → create .agent file → full-screen Copilot experience. Limited to 20 files per agent.
- Microsoft Agent Workspace (experimental, Windows Insider only) — Separate Windows session with own agent account/permissions. Not yet shipping broadly.
- Box MCP Server (live, 2026) — Enterprise content access for agents. Integrated with Claude, Copilot Studio, Azure API Center, Mistral.
- CrowdStrike acquired SGNL ($740M, Jan 2026) — Agent identity/authorization. Signals massive enterprise demand for agent security.
- NIST AI Agent Standards Initiative (Feb 2026) — Both a concept paper on agent identity/authorization AND a formal standards initiative for interoperable, secure agent systems. Public comment deadline: April 2, 2026.
- Gartner — Forecasts agentic AI spending overtakes chatbot spending by 2027 (119% CAGR).
8. Missing Data (Gaps in Available Research)
Section titled “8. Missing Data (Gaps in Available Research)”| What We Don’t Know | Why It Matters |
|---|---|
| Agent file storage specific market sizing | No analyst firm has sized this segment — our TAM is estimated, not sourced |
| Agent-human collaboration market sizing | No one has defined this category yet — we’d be creating it |
| Real MCP server usage data (not downloads) | Download counts are a poor proxy for actual adoption |
| Developer survey data citing file storage as a top pain point | File storage doesn’t appear in top-5 barriers in LangChain or PwC surveys |
| Human demand for agent oversight dashboards | We assume humans want to see agent work — no survey confirms this is a top-3 need |
| Agent infrastructure startup failure rates | The Gartner 40% cancellation prediction is the closest proxy |
| Enterprise spending on agent tooling (not total AI) | The $1.3T IDC number is total AI spending — the agent infrastructure slice is much smaller |
| EU AI Act enforcement specifics for agent workspaces | Aug 2026 deadline — unclear how “human oversight” requirements map to agent workspace features |