LangChain’s $300 million Series C, reported this week at a $3 billion valuation, is the clearest signal yet that enterprises have decided agent tooling is infrastructure rather than an experiment. The company’s paid platform, LangSmith, sells observability: tracing, testing and monitoring for systems that take multi-step actions.
That demand is real, and it also exposes an awkward gap. Most organizations can now deploy an agent. Far fewer can say with confidence whether it is performing correctly across the long tail of tasks it encounters in production.
The measurement problem is structural
A traditional software test asserts an expected output. An agent’s output depends on retrieved context, tool responses, model sampling and the order in which decisions were made. Two runs of the same prompt can be defensible and different.
Evaluation suites help, but they are expensive to build and tend to drift from production traffic. Teams end up with a dashboard of green checks that does not correspond to the failure modes users report.
Procurement is ahead of practice
The consequence is visible in buying behavior. Platform contracts are being signed on the strength of demos and reference customers, with evaluation maturity deferred to a later phase. That is a reasonable way to avoid being left behind, and a poor way to avoid compounding automation errors in finance, support and compliance workflows.
The teams handling this well tend to share a pattern: human review stays in the loop on high-consequence actions, logs are treated as a product rather than exhaust, and agents are scoped narrowly enough that a wrong answer is recoverable.
A market that depends on the answer
If evaluation tooling matures, agent adoption continues and the infrastructure layer compounds. If it does not, the first serious incident at a large customer will slow every deployment in the category. The valuations assume the first outcome. The engineering is still deciding it.

