- Plan → Design → Build → Review: The agent turns a feature request into a task plan and a design spec, implements it, and sends it to code review. Review either merges the change or sends it back to implementation with changes requested.
- Explore → Plan → Code → Commit: The agent reads the repo first, then writes a plan and waits for a human to approve it before editing any files. Approved plans become code and a commit or PR; rejected plans go back to be revised.
- Agent loop (gather → act → verify): The core ReAct cycle behind most coding agents. The agent gathers context, thinks, calls a tool to edit or run something, and checks the result. It repeats until the check passes, then answers.
- Test-driven (red → green → refactor): The agent writes a test from the spec and confirms that it fails, then writes code until the tests pass. The failing test gives the agent a clear, objective finish line, and it cleans up with a refactor at the end.
- Debug (reproduce → diagnose → fix): The agent reproduces the bug before changing anything, finds the root cause, and applies a fix. If verification shows it was the wrong cause, the agent goes back to diagnosis. Once the fix is confirmed, it ships with a regression test.
- Orchestrator + parallel subagents: A lead agent splits a large task among subagents (frontend, backend, tests) that work in parallel. Their results are merged, and a reviewer agent either approves the final PR or sends the gaps it found back to the orchestrator.
- Issue → PR (background agent): The agent picks up an issue ticket, spins up an isolated sandbox, and works on it without a human watching. It keeps iterating until CI passes, then opens a PR for a human to review.
← All posts