Council Multi-Agent System

Challenge

Code review is one of the most valuable quality practices, but it's time-consuming and inconsistent. Human reviewers have different standards, miss different things, and can't always review every PR thoroughly. The challenge was to build a multi-agent system where specialised AI agents could each review code from their domain of expertise — security, architecture, performance, code quality, documentation — and synthesise their findings into a cohesive, actionable report.

Approach

I designed a modular agent architecture with a central orchestrator agent and 6 specialist council members, each with deeply tuned system prompts. The orchestrator analyses the incoming request, determines which council members are relevant, delegates tasks in parallel, and synthesises the results. Each council member has: - A role-specific system prompt with expertise boundaries - Permission restrictions (read-only for reviewers, ask-before-edit for suggestions) - Structured output format with severity levels and line-number references The system uses OpenCode's existing agent framework, extending it with a council skill that provides shared context, review templates, and cross-referencing capabilities.

Key Code

# Council orchestrator delegates to specialists
def delegate_review(code_context):
    members = select_members(code_context)
    reviews = parallel_map(members, lambda m:
        m.review(code_context)
    )
    return synthesize(reviews)

# Each member returns structured feedback
class CouncilMember:
    def review(self, context):
        return {
            "findings": self.analyze(context),
            "severity": self.assess_severity(),
            "suggestions": self.recommend_fixes(),
            "confidence": self.rate_confidence()
        }

Results

• 7 specialised agents providing comprehensive multi-perspective reviews
• Parallel execution reduces review time from minutes to seconds
• Structured output with severity levels, line references, and fix suggestions
• Modular design allows adding new council members without changing the orchestrator

Key Learnings

• Parallel agent execution requires careful prompt isolation to prevent context bleed
• A skilled orchestrator prompt is more important than individual member tuning
• Permission boundaries (read vs write) are critical for safety in autonomous code review

Gallery