LangChain: 89% teams adopt observability

LangChain’s 2026 survey data gives one of the clearest snapshots yet of how production agent teams are instrumenting their systems: nearly 89% of respondents said they have implemented observability for their agents. The company published that figure in its “State of Agent Engineering” report, based on a survey of more than 1,300 professionals including engineers, product managers, business leaders and executives. That number matters because LangChain’s own benchmark places observability ahead of other reliability practices in adoption. The same report says observability outpaces evals adoption, which LangChain put at 52%, suggesting teams are first trying to see what agents are doing before they standardize how to score them. ### Why are teams instrumenting agents so early? LangChain describes agent engineering as the process of turning large language models into reliable systems, and its survey frames observability as part of that core work rather than a later optimization. (langchain.com) The company says organizations entering 2026 are focused less on whether to build agents and more on how to deploy them “reliably, efficiently, and at scale.” LangChain’s production-monitoring guidance makes the reason more concrete. The company says agents differ from conventional software because they rely on model outputs that can vary with prompt wording and because they operate through multi-step reasoning chains, retrieval steps and tool calls that are difficult to anticipate fully during development. ### What does “observability” mean in the agent context? (langchain.com) LangChain defines agent observability as step-by-step visibility into execution. In practice, that means tracing prompts, model outputs, tool calls, retrieval steps and intermediate decisions so teams can reconstruct what happened in a run instead of only seeing a final answer or an error message. The company’s recent guidance also ties observability directly to evaluation. (langchain.com) LangChain said in its January 2026 newsletter that teams shipping reliable agents use a workflow in which production traces feed test datasets and evaluations, because agent behavior “only emerges at runtime.” ### How does that compare with broader enterprise adoption? McKinsey said in an April 2 article that fewer than 10% of companies have scaled agentic AI, even as interest in the technology rises. (langchain.com) The consulting firm argued that scaling depends on foundations such as modernized data architectures, data quality and operating-model changes. (langchain.com) Taken together, the LangChain and McKinsey figures do not prove causation, but they do show a pattern: teams already running agents in production appear far more likely to have observability in place than the broader market is to have scaled agentic AI successfully. That inference is supported by LangChain’s survey of active builders and McKinsey’s reporting on enterprise-scale deployment barriers. (mckinsey.com) ### Why does this push observability beyond debugging? LangChain’s observability materials focus first on debugging and improvement, but the structure of agent traces makes them useful for more than root-cause analysis. A trace can show which tool was called, what context was retrieved, what output was generated and where a workflow diverged, which also makes it usable for operational review and governance records. (langchain.com) McKinsey’s recent infrastructure research points in a similar direction from the enterprise side. The firm said agentic AI infrastructure is becoming the backbone for systems that orchestrate, govern and scale work across enterprises, placing governance alongside execution rather than outside it. ### What should readers watch next? (langchain.com) LangChain’s “State of Agent Engineering” report remains the primary source for the 89% figure, and the company has separately published articles on agent observability and production monitoring. McKinsey’s April 2 and late-April articles remain current reference points for how few organizations have scaled agentic AI and what infrastructure they say is required to do it. (langchain.com) (mckinsey.com)