The consilium Skill — Multi-Agent Panels for High-Stakes Decisions

Most technical decisions go wrong not because the chosen approach was bad in isolation, but because one perspective dominated and the others were never heard. A security analyst would have flagged the unauthenticated endpoint. An SRE would have pointed out the operational complexity. Neither was in the room. The consilium skill simulates that room — a structured panel of specialized agents who evaluate a question from different angles before a decision is locked in.

What it does

The skill runs through six stages: scope, assemble, analyze, debate, converge, deliver. The scope stage forces the decision into a single sentence and classifies the blast radius — how far the decision reaches, from a single module up to business-critical flows affecting revenue or compliance. Blast radius one or two rarely needs Consilium; the skill is designed for three and four.

The assemble stage selects agents from a roster that includes architect, strategist, security, SRE, devops, developer, reviewer, and writer. Preset panels exist for common decision types: architecture decisions get architect, strategist, security, SRE, and devops; production readiness checks get SRE, devops, and security. Custom panels are allowed; the minimum is three agents, the maximum is seven.

Each agent then analyzes the question independently, stating a position (support, oppose, or conditional support) and the specific observations only that agent's role would surface. A security agent names a specific unauthenticated endpoint — not "there's a security risk." After individual analysis, conflicts between agents are surfaced in a debate table with resolution paths. When agents agree after analysis, the debate stage is skipped.

The converge stage resolves remaining conflicts using a priority ladder: safety, correctness, security, reliability, simplicity, cost, elegance — in that order. The deliver stage produces a recommendation with a failure-mode table and a list of conditions that must remain true for the recommendation to hold.

One concrete example

The question: should a new authentication flow use a third-party OIDC provider or a custom JWT implementation? The panel includes architect, security, and SRE. The architect supports the third-party provider for standards compliance. Security supports it for reduced custom-code attack surface. SRE raises a conditional concern about vendor availability and requires a fallback plan. The conflict is logged; the resolution is to add a graceful-degradation path before the decision is finalized. The final recommendation reaches the operator with that condition explicit, not buried.

When it loads

The skill is invoked during /dr-design for Level 3–4 tasks and during /dr-plan for critical decisions with blast radius three or four. It is not invoked for simple bug fixes, Level 1–2 tasks, or questions with clear answers. Over-applying Consilium to trivial decisions wastes time without adding signal.

See the broader context in what Datarim is, or read how design-phase decisions feed into the /dr-design command.