The Gap Was the Product

Tony Kipkemboi wrote a sharp thread arguing that agent frameworks — LangChain, CrewAI, AutoGen, and the rest — are getting squeezed from above by AI labs building their own orchestration and from below by automation platforms bolting on reasoning. His thesis: these frameworks emerged in the gap between when models got good enough and when the infrastructure caught up, and that gap is closing.

He's mostly right, and the implications are uncomfortable.

The honest version of what most agent frameworks do is workflow orchestration. Define tasks, chain them, route data, add conditional logic, call APIs. We've been doing this since Zapier was a Y Combinator demo. The new bit was that the orchestration target — an LLM — was novel enough that existing tools hadn't adapted. So new tools filled the gap.

Two years is not a moat. It's a head start.

Now Claude ships desktop connectors and agent skills. OpenAI launches Frontier as an enterprise agent platform. Zapier and Make add reasoning layers on top of a decade's worth of integrations. The frameworks are caught in the middle, and the middle is thinning.

Kipkemboi's best observation is about who actually sticks with these frameworks long-term: consultancies. System integrators who white-label agentic transformation for clients. That's a real business, but it's not the venture-scale platform play that got pitched in 2023. When your most loyal customers are intermediaries who'll swap you out the moment something better ships, you're a tool, not a platform.

Where I'd push back: the piece assumes AI labs will execute well on the platform layer. History is less generous on this point. Google has been shipping AI capabilities for years without dominating orchestration. OpenAI's enterprise tooling track record is young and unproven. The labs want to own the full stack. Wanting it and delivering it reliably at enterprise scale are different conversations entirely.

There's also a quieter point the piece doesn't quite make. The orchestration problem doesn't vanish when the current crop of solutions gets outflanked — it moves. Infrastructure categories get compressed and re-emerge at a different layer of abstraction all the time. The interesting question isn't whether LangChain survives at its current valuation. It's where the next coordination problem surfaces that neither labs nor automation platforms are positioned to solve.

But the core read is correct. Agent frameworks were an infrastructure response to a temporary gap in the market. The gap is closing. What remains is open-source community projects and niche developer tools — both viable, neither venture-scale.

The smartest founders in this space already know. The ones still pitching "the orchestration layer for the agentic era" are selling a map of a country that's being annexed from both borders.