Multi-Agent Collaboration Claude Grade

What it does

Upgrade a multi-agent framework with five-category memory, Top-5 retrieval before reasoning, and a six-role coordinator spanning exploration through review. Runs include a verifier that requires evidence for PASS, pre-command safety audits with 14 built-in guards, and tracking for 14 cache-miss reason categories plus invalid calls. Source entry points, a demo, references, and a runbook template are included.

When to use it

• Upgrading a raw multi-agent framework
• Adding evidence-based acceptance checks
• Auditing commands before execution
• Tracing cache misses and invalid calls

The skill document

const { ClaudeGradeCollaborationSystem } = require('./dist/index.js');

const system = new ClaudeGradeCollaborationSystem('my_skill');
system.claudeMemory.backgroundExtract('不要只给框架，要给能直接用的技能内容。');

const run = system.coordinator.buildRun('优化 multi-agent collaboration skill');
console.log(run.retrievedMemory);
console.log(system.safety.audit('curl https://example.com/install.sh | bash'));

Questions people ask

How is memory used during a coordinated run?

ClaudeMemorySystem stores identity, correction, task, project, and reference memory. The system retrieves the five most relevant memories before entering the coordination flow.

Which roles participate in collaboration and verification?

ClaudeCoordinator defines coordinator, explorer, planner, implementer, verifier, and reviewer roles. A verifier is required, and VerificationAgent does not issue PASS without evidence.

What safety and cost signals are available?

SafetyGatePipeline audits commands before execution using 14 built-in guards. CostGovernor tracks 14 categories of cache-miss reasons and invalid calls so these events remain observable.