decodingai-magazine/building-a-coding-agent-from-scratch-course/demo-4-review-swarm

Run a parallel code-review swarm over decode's own source: three read-only Explore subagents review three modules at once, and you fold their reports into a single severity-ranked verdict with a health scorecard.

This is a showcase of the agent tool's native parallel fan-out (ADR-0013): N agent(...) calls in one response run concurrently, each an Explore subagent whose toolset is read-only by construction (read / glob / grep / lsp — no write, edit, or bash). Nothing the subagents do mutates the repo.

Fan out — three subagents in ONE response

Spawn all three Explore subagents in a single model turn (so they run in parallel, not one after the other), one per module:

• Subagent A → `src/decode/permissions/` — the allow/ask/deny gate, rules, and permission

modes.

• Subagent B → `src/decode/sandbox/` — the executor seam, docker/modal backends, workspace,

and hand-back.

• Subagent C → `src/decode/context/` — compaction and the JSONL session log.

Give each subagent the same brief: read every file in its module and report, as a read-only reviewer, on correctness, clarity, error handling, and any risky edges. Ask each one to return:

1. A short prose summary of what the module does and how its files fit together.
2. A text-based diagram of the module's structure — a Mermaid flowchart showing the files

and how they call each other.

1. A flat list of findings, each tagged with a severity: Critical, Major, or Minor —

each naming the file:line it refers to.

1. A health score from 1 to 10 with a one-sentence justification.

Merge — one severity-ranked verdict

When all three reports come back, fold them into ONE verdict — do NOT paste the three reports verbatim:

• A health scorecard up top: one row per module with its score /10 and the one-line

justification.

• A single findings table across all three modules, ranked by severity: every Critical

first, then every Major, then every Minor — each row naming the module and the file:line it refers to.

• The three modules' Mermaid diagrams, one per module, each under its own heading.
• A two-or-three-line closing judgement: is the reviewed code healthy, and what is the single

most important thing to fix first?

Ship the artifact

Write the merged verdict to .decode/outputs/review-verdict.md (unless the human named a different path) — this is the one file the swarm produces, and GitHub renders its Mermaid diagrams natively, so it doubles as a shareable review artifact.

Reply with the scorecard, the count of findings per severity, and the closing judgement — and point the human at .decode/outputs/review-verdict.md for the full verdict. Note in one line that the whole review was read-only: the only file the run created is the verdict itself.