Decomposing work for AI agents

AI agents write better code when the work is decomposed clearly. Vague epics produce vague implementations. The decomposition standard — design to feature to epic to story to task — is what makes agent output reviewable.

The decomposition chain

Each level answers a different question:

| Level | Question | AI role | |-------|----------|---------| | Design | What problem and constraints? | Draft, human approves | | Feature / Epic | What capability ships? | Decompose, human prioritizes | | Story | What user-visible behavior? | Expand, human refines | | Task | What technical work? | Break down, human assigns | | Implementation | What code changes? | Agent drafts, human merges |

Skip a level and review becomes guesswork.

Where human gates belong

Not every stage should be fully automated:

• Design and epic boundaries — human decision. AI proposes; leadership approves trade-offs.
• Story acceptance criteria — human-written or human-edited. Agents need a rubric.
• Task-level implementation — agent-assisted, with eval before merge.
• Production deploy — human approval. Always.

Eval before merge

Agent-generated PRs need the same quality bar as human PRs — defined explicitly:

• Lint and test pass rate
• Review rejection criteria (what sends it back)
• Scope check (did the PR do only what the task specified?)

Without eval, you measure AI by volume. With eval, you measure it by throughput and quality.

Starting point

Pick one feature currently in progress. Decompose it through all five levels on paper. Then run the agent on a single task with that rubric. Compare review time and rejection rate to your baseline.

That is where AI in the SDLC begins — not with a license purchase.

For help building decomposition standards and eval rubrics, see AI in the SDLC.