1. From autocomplete to agents changes the bottleneck
Tools like Claude Code, Cursor in agent mode and GitHub Copilot agent mode do not just suggest a snippet. They read the repository, plan, edit multiple files, run tests and open a pull request with a convincing summary. The bottleneck moved from writing code to reviewing, integrating and maintaining what the agent produced.
The Stack Overflow Developer Survey 2025 shows high adoption alongside distrust of accuracy, and the DORA 2025 report reinforces that AI amplifies the maturity a team already has. Teams with tests, review and predictable deploys gain more; disorganized teams just accelerate instability, now in larger diffs.
- Treat a generated pull request as untrusted until fully reviewed.
- Prefer reviewing the plan before letting the agent execute everything.
- Keep the diff small enough that review is actually possible.
- Do not paste secrets, tokens or private customer data.
- Record important decisions instead of relying on chat history.
2. Specs and project instructions matter more, not less
An agent executes what you specify. A weak spec makes it build the wrong thing quickly, across many files, with tests that confirm the mistake. Describing the goal, constraints, relevant files, acceptance criteria and what must not change is now the cheapest way to avoid rework.
Beyond each task spec, a project instruction file such as AGENTS.md or CLAUDE.md gives the agent permanent context: build and test commands, conventions, architecture and what to avoid.
- Provide exact error messages, relevant files and existing patterns.
- State technology versions and what should not change.
- Explain business rules and edge cases the agent cannot infer.
- Ask for a plan and a small change before broad rewrites.
- Keep an AGENTS.md or CLAUDE.md with commands, conventions and limits.
3. Review the whole diff, and distrust generated tests
AI can produce code that compiles and passes the tests it wrote itself while violating security, performance, data privacy or an undocumented business rule. The agent summary describes intent, not necessarily what changed, so read every file, including the ones touched in passing.
Generated tests often confirm the implementation rather than the expected behavior. In a bug fix, require at least one test that would fail without the change.
- Read new dependencies, config, Dockerfile and migration changes in the diff.
- Require a test that fails without the fix and passes with it.
- Review database queries, permissions and data exposure.
- Limit what the agent can run and access in environments with real data.
- Use pull request review even when AI writes most of the patch.
4. Measure throughput and stability, not generated lines
Counting lines or pull requests generated by AI says nothing about value. DORA measures delivery and stability together: deploy frequency, lead time, change failure rate and recovery time.
If the agent raises speed but worsens the failure rate, the gain is an illusion. Prompt injection and excessive agency are real risks once an agent reads issues and logs and runs commands, as the OWASP Top 10 for LLM Applications 2025 describes.
- Track change failure rate, not just how much code was generated.
- Do not treat content the agent reads as trusted instructions.
- Keep final approval tied to team standards.
- Add or update tests for critical behavior before deploy.
- Plan rollback for changes that affect sensitive flows.
How to use this article
Treat this page as a decision aid. Use it with the related hub, checklist or service route when the topic affects production, customer experience, deployment, security or business continuity.