Stack Overflow Pulse Survey: AI agents are already in the works, but not yet autonomous (Category AI4SDLC)
Dealing with the new one. Stack Overflow Pulse Survey on Agentic AI The study came out 27 May 2026 year and based on survey 1 100 Developers held at the end of April. The main result is that agents have become more popular, but the popularity is growing not through complete autonomy, but through a rather tough human control. (human in the loop).
The share of respondents who use AI-agents at work at any frequency has increased. 31Percentage of Developer Survey 2025 before 59percentage results Big StackOverflow survey 2025 years. Added a picture from the report, where the trend is very noticeable.
Really, 63The percentage of respondents rarely or never allow agents to operate fully autonomously. 60The percentage blocks unapproved system changes. That is, the real picture does not yet look like the agent does the work from beginning to end - the agent starts the work, but the rights to act remain limited. He can write, suggest, search, explain, collect a draft solution. But when the state of the system begins to change, a person appears, approval, diff, CI, policy or other gate. And that fits well with the verification cycle trend I mentioned earlier.
I wonder what. 68Percentage of respondents prefer predictable single-agent configurations over complex multi-agent systems. Among the current agent workers of the protuesses 69Percentage of respondents use one agent (illustrated 69%, and in the Tex study 68Percentage - who to trust:) ), 17% - several specialized agents, 16Percentage: Several intersecting or coordinated agents. In general, the market likes beautiful swarm schemes, but practitioners still choose one agent, a clear task, a visible result, and less coordination magic.
Daily use is noted 40% full-stack developers, 52% of architects and 50% of top managers (Senior executives in the StackOverflow survey are usually engineering leaders.). But it's the audience's self-assessment of the Stack Overflow pulse survey, not general industry information. The most enduring barriers are also expected: accuracy and safety. Spending has become a less painful constraint: the share of those who consider cost a barrier has decreased from the previous year. 53Percentage of Developer Survey 2025 before 38% in the pulse survey. But accuracy and safety remain the main risks, although the share of strongly agreeing also decreased. 57percentage 47% for safety with 56percentage 44%.
Interestingly, AI is useful enough to become a de facto standard, but not reliable enough to remove human-in-the-loop. Money becomes an operational issue and quality, safety and responsibility becomes an architectural one. The instrumental landscape is recognizable: GitHub Copilot, Claude Code, OpenAI Codex, Cursor, Replit, Lovable, v0, LangChain, LangGraph, OpenAI Agents SDK. But for me, the most interesting interest in observability/evals tooling is that users are more likely to want to expand the use of such tools than they are already actively using them. If an agent becomes part of a production development system, it needs more than just prompts and access to the repository. You need rights, logs, evals, audit trails, kickbacks, policies, liability and a clear way to prove that the action was safe.
P.S. StackOverflow research is available on our website ai4sdlc-research.spaceWe will soon relaunch our AI4SDLC research, this time about agent penetration into development processes.
#AI #AI4SDLC #Agents #Engineering #Software #Management