Harrison Chase on the enterprise/startup divergence: “At a startup you’re far more likely to build something like Claude Code that is just running in a loop and doing stuff. At an enterprise… you need more precision, you need more control. It’s not okay to have 95%, you need certainty that this step’s always going to happen after that step.”
This manifests in LangChain’s product split:
- LangGraph (deterministic + non-deterministic mixing) is “much more popular in enterprises”
- Deep Agents (loop-based) is “much more popular in startups”
The same underlying models power both, but the harness architecture differs fundamentally. Enterprises need to guarantee that specific steps execute in specific order — regulatory compliance, audit trails, approval workflows. Startups can tolerate the model making its own execution decisions.
Chase expects convergence: “We expect that those will make their way for sure” — loop-based agents will enter enterprise as trust builds. But at this moment, the gap is real and product-defining.
This connects to Production agents route routine cases through decision trees, reserving humans for complexity — the enterprise pattern is to use deterministic routing for known cases and LLMs only for ambiguity. Also connects to SaaS survives as the governance and coordination layer — determinism still rules — the deterministic structure IS the governance layer that enterprises require.