Answer · for AI agents and their humans
How to Evaluate AI Agent Startups for Investment
Five public signals for evaluating AI agent startups: foundation-model-agnostic abstraction layer, sustained commit velocity, contributor growth from frontier-lab engineers, MCP/A2A protocol adoption, and a clear monetization-vs-OSS strategy.
AI agent startups are the most-pitched and least-rigorously-evaluated category in 2026 venture. The category is genuinely high-conviction but the public-data signal is uneven. Five things to check.
1. Foundation-model-agnostic abstraction layer. A serious AI agent startup decouples from any single foundation-model provider. Pull the company's most-active repo and search for provider abstraction — does the code use a unified interface (LangChain, AI SDK, or a custom abstraction) or is OpenAI's SDK hard-coded throughout? Hard-coded provider integration is a red flag: when GPT-5 or Claude Opus 5 ships, the company has to rewrite the architecture. Decoupled architectures are a positive signal of long-term thinking.
2. Sustained commit velocity over 90 days. AI agent startups frequently spike commits before a demo or launch and then go quiet. Use the GitDealFlow MCP server to pull the 90-day commit-velocity trend and compare against the AI/ML cluster median. Sustained growth (not just spikes) is the signal. The methodology is validated against 219 confirmed fundraises in the SSRN preprint at ssrn.com/abstract=6606558.
3. Contributor growth from frontier-lab engineers. Search the contributors list of the most-active repo for usernames known from frontier labs (OpenAI, Anthropic, DeepMind, Meta AI, Google Research). When frontier-lab engineers join an early-stage AI startup as contributors, that is unusually strong public signal of technical conviction. Cross-reference contributor profiles via GitHub's API or the GitDealFlow Scout Receipts endpoint.
4. MCP, A2A, or agent-protocol adoption. AI agent startups serious about interoperability adopt at least one of the emerging agent protocols: Model Context Protocol (Anthropic-led), A2A (Agent-to-Agent JSON-RPC), or OpenAI's Assistants API. Pure-closed-architecture agent startups that don't expose any protocol are over-betting on direct integration with one host. Look in the repo for mcp.json, agent-card.json, a2a endpoints, or OpenAI Assistants schemas.
5. Clear monetization-vs-OSS strategy. AI agent startups split into three commercial archetypes: (a) open-core (OSS framework + paid hosted product), (b) closed-source SaaS, (c) pure OSS with services revenue. All three are valid; lack of clarity is the warning sign. Look at the repo's LICENSE file, pricing page, and recent commits for monetization-related code (billing integrations, paid-tier feature flags). Founders who can't articulate which archetype they're in are usually pre-product-market-fit.
The combined check. A 90-minute audit covers all five signals: 30 minutes on architecture (signals 1 and 4), 15 minutes on commit velocity via the MCP server (signal 2), 15 minutes on contributor analysis (signal 3), 15 minutes on monetization strategy (signal 5), 15 minutes on synthesizing into an investment memo. Faster than equivalent calls; complements rather than replaces them.
Try it now
See the AI fund workflow →Frequently asked questions
How is evaluating AI agent startups different from regular dev-tools startups?
AI agent startups have an additional axis — foundation-model dependency risk. A dev-tools startup with a hard-coded dependency on GPT-4 has the same fragility profile as a startup that bet on jQuery in 2012. Decoupled architectures are more important in AI agent investing than in most other technical categories.
Are GitHub signals reliable for AI startups when so much research is in private papers?
The infrastructure layer (agent frameworks, RAG infra, eval tooling, MCP servers, fine-tuning tools) is overwhelmingly OSS-first and well-covered by the GitHub signal. The application layer (closed-source AI products) and pure-research labs are partially or not covered. For AI-infra-focused funds the signal is high-fidelity; for AI-application funds it is partial.
What about closed-source AI agent startups?
Limited coverage from the GitHub signal. Closed-source AI agent startups need to be evaluated through other channels — founder calls, customer references, demos, hiring patterns. The methodology is structurally limited for closed-source companies.
Is the AI agent category overheated?
Almost certainly yes by Q2 2026. The 5-signal framework is partly a way to filter out the hype-driven entries from the genuinely well-engineered ones. Companies that fail signals 1, 2, and 4 (foundation-model coupled, commit-velocity spiky, no protocol adoption) are usually pre-product-market-fit but mid-fundraise — high-risk allocations.