AI Coding Agents: How Software Teams Use Them Well

Coding agents changed faster than any other part of the AI stack in 2025 and 2026. This post collects what the large surveys show about how teams use them, where the evidence is weaker than the marketing, and the working practices I recommend to clients adopting them.

Adoption: near universal, now agent-shaped

Google's DORA 2025 report, a survey of nearly 5,000 technology professionals, found AI adoption among software development professionals reached 90%, with a median of two hours a day spent working with AI.

The type of tool has changed too. In Stack Overflow's 2025 survey, about 31% of developers used AI agents at work in some form, and 37.9% did not plan to. By Stack Overflow's April 2026 pulse survey, agent use had almost doubled to 59%.

The full 2026 Developer Survey found that 66% of respondents use coding assistants or agents, and 73% of those use them daily. Among the coding agents respondents reported using, Claude Code (66%) and GitHub Copilot (59%) led.

The main tools, by working style

The tools have converged on features, so the useful distinction is how each fits into a developer's day.

Stack Overflow's 2026 page does not report a figure for Cursor, so I have not put a usage number next to it. Many teams use more than one: an editor tool for small edits and a repository-level agent for larger tasks.

What developers actually use them for

The 2026 survey shows what tasks dominate (source):

• generating code in a familiar area: 69%,
• debugging, troubleshooting or refactoring: 64%,
• answering straightforward technical questions: 59%,
• writing or improving tests: 58%,
• generating code in an unfamiliar area: 56%.

Stack Overflow's results write-up adds that only 20% use AI for deploying, operating or troubleshooting production systems. Agents are trusted with code that gets reviewed, much less with live systems. That is the right instinct.

Productivity: real, uneven, hard to measure

Self-reported gains are strong. In DORA 2025, over 80% of respondents said AI had improved their productivity, and 59% reported a positive effect on code quality. Among Stack Overflow's 2025 agent users, about 69% agreed agents had increased their productivity.

Controlled measurement is more cautious. METR's randomised study of experienced open-source developers in early 2025 found tasks took 19% longer when AI was allowed, even though developers believed it had sped them up by 20%. METR's February 2026 follow-up with late-2025 tools estimated a speedup instead (18% for returning developers, 4% for new ones), but with wide confidence intervals and selection bias: many developers no longer wanted to work without AI at all. METR's own conclusion is that developers are probably faster now, and the evidence on how much is weak.

DORA's framing is the most useful one I have seen: AI acts as an amplifier, magnifying the strengths of well-run teams and the problems of struggling ones. A team with good tests, clear architecture and fast review gets more out of agents. A team without those gets more code it cannot trust.

Trust is the bottleneck

Developers have not become less sceptical. In 2025, 66% said their top frustration was AI output that is "almost right, but not quite". In DORA 2025, only 24% reported a great deal or a lot of trust in AI. And in 2026, 48% said they trust AI when they can easily validate the answer.

"When they can easily validate" is the key phrase. Agents are most useful where verification is cheap: a test suite, a type checker, a preview deployment, a reviewer who knows the code.

Cost is becoming a management issue

Agents consume far more tokens than autocomplete, and vendors are moving from seats to usage. Gartner predicts AI coding costs will surpass the average developer's salary by 2028 as consumption grows. Stack Overflow's 2026 results show organisations reacting: 38% set usage limits and 28% review spending data.

How to adopt coding agents well

Give agents context, not just prompts

Agents work from what they can read. Keep a short project guide in the repository (architecture, conventions, commands to run tests), and expose internal knowledge through tools. My design system MCP server is an example: it gives coding agents the real components and tokens so they build on-brand UI instead of guessing at styles.

Make verification cheap

Tests, types, linters and preview deployments are what turn "almost right" into "checked." Teams with weak tests should invest there before scaling agent use.

Keep pull requests small and reviewed

Ask agents for focused changes. Large AI-generated diffs are hard to review, which is exactly where subtle bugs get through. A human approves every merge.

Keep agents away from production by default

Read-only access to logs is useful. Write access to production systems, databases or secrets should require explicit approval, and most teams should not grant it at all.

Set budgets and measure

Track cost per developer and compare it with delivery metrics you already trust: cycle time, change failure rate, escaped defects. Self-reported speed is not enough, as the METR results show.

Do not stop developing people

Junior developers need to understand the code an agent writes. Use agents to explain and to review, not only to generate.

Work with me

I build full-stack and AI products with modern AI-assisted workflows, and I set up the context, tooling and review practices that make coding agents useful for a team. See full-stack development and MCP server development, or book a scoping call.