PulseFlow Tecnologia
← Insights
·6 min read·product-update·langchain·state-of-ai-agents·data
PulseFlow Tecnologia

LangChain State of AI Agents: what the data shows about agents in production

More than 1,300 professionals responded to LangChain's State of AI Agents survey - 60% from the tech sector, 51% from startups and small companies with fewer than 100 employees. The result is the most detailed portrait available of where agent adoption actually stands, not where the hype suggests it should be.

Adoption is already the majority, not the minority

51% of surveyed companies already have agents in production, and 78% have active plans to expand. Mid-sized companies (100 to 2,000 employees) lead, with 63% already in production - and adoption is nearly identical between the tech sector (89%) and other sectors (90%), which dismantles the idea that only tech companies are taking this seriously.

The three use cases that dominate

Research and summarization leads with 58% adoption, followed by personal productivity/assistance (53.5%) and customer support (45.8%). These are tasks where input and output text matters, but that still depend heavily on human judgment in review - none of the three is yet dominated by full autonomy.

What actually blocks adoption

Performance quality is the top reported concern, weighted twice as heavily as any other factor - not security, not cost, execution quality itself. Security shows up as critical specifically for large companies. Cost is cited by 22.4% of small companies as a barrier, a smaller number than you'd expect, suggesting the technical problem (the answer isn't good enough) weighs more than the financial one at this stage.

The controls that separate mature adoption from risky adoption

Tracing and observability tools are the most-cited control among surveyed companies. Offline evaluation (39.8%) already outpaces online evaluation (32.5%) - a sign teams prefer catching problems before production, not after. Read-only permissions dominate most cases, with human approval reserved for critical actions like writes or deletes. The most revealing difference: tech companies use 51% more simultaneous control methods than companies in other sectors - and not by coincidence, they're the ones reporting the least friction in adoption.

The most-mentioned applications

Cursor (AI-powered code editor), Perplexity (AI-powered search engine), and Replit (accelerated development) show up as the most-discussed applications among respondents - all tied to technical productivity, reinforcing that software engineering remains the most mature use case for AI agents.

Why this report matters for anyone deciding on adoption

The most actionable data point isn't "51% already adopted" - it's that tech teams didn't adopt faster because they had access to a better model; they adopted faster because they implemented more simultaneous controls. That inverts the intuitive logic that control slows down speed - for AI agents, control is what enables sustainable speed.

Sources