Agent Arch Board

Agent architecture board

An agent that proposes and blesses its own design produces confident nonsense with nothing to check it. A board separates the roles: a proposer argues for a design, several reviewers judge it blind to each other, and a chair reconciles the verdicts into one binding decision. The value lives in the independence, so reviewers who can read each other's scores converge and stop thinking.

Method

1. Proposer writes the design. Load principal-architect-role. Output rfc.md in the shape the rfc-process skill defines: context, at least two options weighed, the recommendation, and the trade-offs it accepts. A proposal with one option is advocacy, not design.
2. Fan out to blind reviewers. Spawn reviewers in parallel, each in an isolated context that cannot see the others. Give each a lens: load cloud-architect-role for cost and scale, security-engineer-role for threat surface, site-reliability-engineer for operability. Blindness is enforced, not requested.
3. Each reviewer scores against a rubric. Output review-<lens>.md: a verdict of approve, approve-with-conditions, or reject, with rationale tied to named RFC sections. A reject must state what would turn it into an approve.
4. Chair synthesizes a binding decision. Load staff-engineer. Read every review together, reconcile the conflicts, and issue decision.md as an architecture decision record: the call, the conditions attached, and which review points drove it. The chair may overrule a reviewer but must say why in writing.
5. Split verdicts escalate, they do not average. When reviewers land on opposite sides of a point neither will concede, the chair sends it to a human architect rather than splitting the difference. Averaging two incompatible designs yields one that serves neither.
6. Conditions are tracked to closure. Every approve-with-conditions item becomes a checklist the proposer clears before the design ratifies. An open condition keeps the decision provisional.

Run it

In Claude Code, run the proposer as one subagent, then spawn the reviewer subagents in a single parallel batch so none sees another's output, then run the chair over the collected reviews. Keep rfc.md, the review files, and decision.md as artifacts on a design branch, mirroring an architecture-review-board. Terminate when the chair issues a decision with every condition listed, or when a split escalates to a human. To port, use LangGraph with a fan-out-to-reviewers node and a synthesis node, a CrewAI judge panel with parallel review tasks feeding a chair task, or AutoGen reviewers messaging a moderator that withholds cross-talk until submissions close.

Signals it works

• Reviewers submit before seeing each other, so no review anchors on another.
• Every reject names the change that would flip it to approve.
• The chair's decision cites specific review points, not a vote tally.

Boundaries

This board decides technical direction, not budget or staffing, and it defers a genuinely novel bet to the human owners who carry the risk. It assumes the proposer produced a real RFC; garbage in still yields a confident decision out. Ratifying anything irreversible waits on a human sign-off.