LangChain: Large companies are deploying AI agents into production faster than startups

LangChain released the State of Agent Engineering report based on a survey of 1,300 specialists from various companies. The main finding sounds unexpected: AI-agents are already entering real business not where there is less bureaucracy, but where there are more resources and stricter reliability requirements.

Who's already in production

In the report, AI-agents are understood as systems that don't just respond to a request like a regular chatbot, but choose their own steps to achieve a goal: search for data, call tools, send emails, or trigger actions in other services. And "production" in this context is not a demo or pilot, but a live environment where the product is already being used by employees, clients, or partners. This is where the report breaks the popular stereotype that large companies are too slow for new technologies.

According to LangChain, 67% of organizations with more than 10,000 employees have already deployed agent applications to production. Among companies with fewer than 100 employees, this figure is 50%. The reason looks pragmatic: reliable agents require infrastructure, integrations, monitoring, and a team that can maintain all of this — and enterprise businesses usually have more such resources.

Where the weakness lies

The second important part of the report concerns not launching, but quality control. Here it's useful to distinguish between two terms. The first is observability — the team sees what the agent does, which tools it calls, at what step it makes mistakes, and why it came to a specific result. The second is offline evaluation, which is checking against a test set of tasks with known correct answers. This is no longer observation of consequences, but an attempt to measure quality before real-world use.

The gap between these practices turned out to be significant. 89% of surveyed teams have implemented observability mechanisms, but only 52.4% conduct offline evaluation. In other words, the market has learned much better how to watch agent behavior after launch than to systematically check them beforehand. This approach could be described as "deploy first, figure it out later." For regular software this is risky, and for agent systems even more so: an error here can turn into not just incorrect text, but incorrect action.

Why quality became the bottleneck

Another shift concerns money. A few quarters ago, conversations about AI-agents almost always came down to the cost of models and infrastructure. In the new report, the picture is different: 32% of respondents name quality, not price, as the main barrier. By quality here they mean not abstract "let it be smart," but quite practical things that directly affect trust in the system and the business's willingness to expand deployment.

• Accuracy of agent answers and actions
• Stability of results from run to run
• Minimizing hallucinations and false conclusions
• Acceptable latency between request and response
• Security and compliance requirements

The second-most important barrier depends on company size. Startups more often complain about latency — the delay that makes interaction with the agent slow and frustrating. Large companies with more than 2,000 employees more often point to security and compliance. The logic is clear: the larger the business, the higher the cost of error, the stricter the requirements for data, audit, access, and reproducibility of each step.

What this means

The AI-agent market is maturing rapidly. The question is no longer whether you can assemble a beautiful demo scenario, but whether you can stably and safely integrate an agent into a work process. Therefore, the next stage of competition will not follow the line of "who has the cheaper model," but along the line of engineering discipline: who tests better, observes better, and limits errors before the agent starts acting on behalf of the user.