Agent design review panel

Agent design review panel

Ask one reviewer to check "everything" and it checks the first thing it thinks of, then calls the job done. Split the concern into fixed lenses, give each its own agent, and the security hole the correctness reviewer skimmed past gets caught by the agent that looks for nothing else. The shape is a parallel judge panel: reviewers run independently, then a moderator merges their verdicts.

Team

• Correctness (code-review): does the design meet the spec on every path.
• Security (security-engineer-role): injection, authz gaps, secret handling.
• Cost (cloud-architect-role): compute, latency, and spend at real scale.
• Simplicity (principal-architect-role): what can be deleted and still work.
• Moderator (staff-engineer): merges the four verdicts into one decision.

Method

1. Give every reviewer the identical packet. The design doc, the spec it answers, and the constraints. Same input, different lens, so a disagreement reflects the lens and not a difference in what each one read.
2. Run them in isolation. No reviewer sees another's findings before forming its own. Cross-talk collapses four independent opinions into one loud one, which is the exact failure the panel exists to prevent.
3. Fix a common finding format. Each reviewer returns findings as severity, location, concern, recommendation plus a one-word lens verdict: approve, revise, or reject. A uniform shape lets the moderator merge without re-reading four styles.
4. Merge, dedupe, and rank by lens priority. The moderator collapses duplicate findings and resolves conflicts by a stated order: a security reject outranks a cost revise. It never overrules a lens on that lens's own turf.
5. Emit one verdict with the blocking lens named. Approve only when no lens rejects. On a revise, the author fixes and the panel re-runs, capped at two rounds before a human breaks the tie.

Run it

In Claude Code, spawn one subagent per lens in a single parallel Task batch, each prompt carrying its lens and the shared packet, and forbid them from reading each other's output. A separate moderator subagent reads the four finding lists and writes the synthesis. Store each review and the synthesis as files so the author sees who blocked and why. Terminate on a clean approve across all lenses, or on the revise cap with escalation to a human. To port, this is a CrewAI parallel task group feeding a synthesis task, an AutoGen GroupChat with a summarizing manager, or a LangGraph fan-out to reviewer nodes joining at a reducer.

Signals it works

• Findings cluster by lens, and no two reviewers report the same issue.
• The verdict names the exact lens that blocked, or approves cleanly.
• Reviewers reach different conclusions, proof they did not anchor on one voice.

Boundaries

A panel reviews a design or plan, not a line-level diff: for author-reviewer iteration on code use agent-code-review-loop. Four lenses is a default, not a law, add a performance or accessibility lens when the artifact demands it. The moderator arbitrates priority; it does not invent findings the reviewers missed.