Agentic SDLC workflow

Agentic phases

1. Investigate: inputs include the story brief, codebase context, and constraints. Outputs are findings, assumptions, risks, and impacted components.
2. Plan: outputs include the step-by-step plan, file list, test plan, and rollout or rollback considerations.
3. Implement: outputs include incremental changes with checkpoints such as build, compile, or tests.
4. Self-review: outputs include review notes, checklist pass or fail, and suggested improvements.
5. PR packaging: outputs include PR description, validation evidence, and risk notes.
6. Review iteration: outputs include targeted fixes responding to feedback with minimal scope expansion.

Ownership model

• Agent: proposes and drafts.
• Engineer: validates, decides, and owns final output.
• Reviewer: enforces standards and checks for subtle drift.

Quality gates

• No plan means no code.
• No tests means no merge unless an exception is explicitly documented.
• Parity evidence is required for migrations.
• PR size and scope limits reduce review risk.
• High uncertainty triggers a stop condition and a human-first analysis path.

When not to use agents

• High-ambiguity requirements without a stable source of truth.
• Deep work in novel libraries where wrong assumptions are costly.
• Security-sensitive changes without an explicit review path.
• Large refactors that lack incremental safety nets.