Coding agent workflow

Coding agent workflow

A coding agent is fast and tireless but confidently wrong when under-directed. Getting reliable work out of one is a skill: brief it tightly, keep changes reviewable, insist on verification, and stay the accountable engineer, not a spectator.

Method

1. Brief with a goal and a done-condition, not a vibe. State exactly what you want, the constraints, and how to know it is complete (tests pass, the behavior works, the output matches). A vague task gets a vague, plausible result. Point the agent at the relevant files and conventions rather than letting it guess (see agent-complete-task for the prompt shape).
2. Keep changes small and reviewable. One task per run; a focused diff you can actually read beats a sprawling change that touches forty files. Large tasks get decomposed into steps you review between (see agent-task-breakdown, pull-request-size's logic applied to agents).
3. Make it explore before it edits. The agent should read the code and match its patterns, not impose a generic style. Instruct it to understand first; unfamiliar-codebase mistakes come from skipping this (see agent-explain-codebase).
4. Demand verification, and check it. Require the agent to run the tests, build, and linter, and to exercise the actual behavior, then report what it ran. Then verify yourself: agents claim success they have not demonstrated. "It should work" is not "it works" (see agentic-loops' verify-before-done rule, verify skill).
5. Review the diff like any code review. Read what changed and why, not just whether it runs. Watch for scope creep (unasked refactors), silent behavior changes, weakened tests, and confident comments over subtly wrong logic. You are accountable for merging it (see review-my-code, code-review).
6. Gate the irreversible. Destructive or outward-facing actions (deleting, force-pushing, deploying, sending) get your confirmation, not the agent's autonomy; keep it working on a branch, and keep the changes reversible (see llm-guardrails, automation-guardrails).

Boundaries

• The agent accelerates the work; it does not transfer accountability. Merging unreviewed agent output because it looked confident is how bad changes ship.
• Agents excel at well-specified, verifiable tasks and struggle with ambiguous ones; the payoff tracks how well you can state the goal and check the result. Underspecified creative architecture is still yours.
• This is human-directs-agent supervision; multi-agent orchestration (agents coordinating agents) is a separate discipline with its own costs (see agent-orchestration-antipatterns, multi-agent-workflow).