Multi-Agent Orchestrator Prover MCP Connector for Claude

You must: (1) define ROLES — each agent has a name, typed input/output contracts, responsibility boundary, and explicit exclusions. "General purpose" is not a role, (2) specify HANDOFFS — each agent-to-agent transition has trigger conditions, typed data contracts, and explicit failure behavior. "Passes data" is not a protocol, (3) contain FAILURES — each agent has timeout, retry policy (count + backoff), fallback, and circuit breaker. Pipeline has cascade protection. "Retry and it works" is not containment, (4) define CONSENSUS — when agents conflict, specify detection method, resolution protocol (voting, scoring, supervisor), and deterministic tie-breaking. "They agree" is hope, not architecture, (5) enforce OBSERVABILITY — correlation IDs per request, spans per agent action, per-agent metrics (latency, tokens, errors), dashboards, and alerting. "Check the logs" is not observability. If rejected, fix the specific orchestration flaw before deploying. Structured reflection tool for multi-agent system design — forces explicit role boundaries, typed handoff protocols, per-agent failure containment, deterministic consensus mechanisms, and distributed observability before deploying any agent pipeline or agentic workflow. Catches Roles Undefined (agents with overlapping or vague responsibilities — a 4-agent pipeline where both the ResearchAgent and AnalysisAgent "summarize relevant information." Result: both agents produce summaries, the WriterAgent receives two partially overlapping inputs, the final response contains contradictions (ResearchAgent found 12 sources, AnalysisAgent found 8, 5 overlapped, 3 conflicted). 23% of pipeline outputs contained factual inconsistencies traceable to role overlap. Each agent must have: a name, explicit input/output contracts with typed fields, responsibility boundaries, and exclusion clauses — "does NOT evaluate quality" is as important as "DOES generate responses"), Handoffs Missing (agent-to-agent transitions that lose data or have no failure behavior — ResearchAgent returns confidence: 0.3 on a query. No handoff rule exists for low confidence. WriterAgent receives the low-confidence research and generates an authoritative-sounding response based on unreliable data. The user sees a fluent answer with no indication that the underlying evidence was weak. A handoff protocol must specify: trigger conditions (confidence > 0.7 AND sources >= 2), typed data contracts (what fields transfer), and failure behavior (retry with expanded query, escalate to human, return partial results with confidence warning)), Failures Cascading (one agent failure kills the entire pipeline — a 5-agent pipeline with no circuit breakers. The ExternalAPIAgent calls a rate-limited API. At 3 AM, the API returns 429 Too Many Requests. The agent retries infinitely. The pipeline hangs. 847 requests queue behind it. By 6 AM, the orchestrator OOM-kills. Every agent needs: timeout (hard limit), retry policy (count + backoff), fallback (cached results, degraded response), and circuit breaker (open after N consecutive failures). The pipeline needs: cascade protection — if >1 agent circuit is open, return graceful degradation), Consensus Absent (multiple agents produce conflicting outputs with no resolution mechanism — a fact-checking pipeline: VerifierAgent says "claim is true (confidence: 0.72)." CounterResearchAgent says "claim is false (confidence: 0.68)." No resolution protocol exists. The system returns... both? Neither? The first one? Random? Consensus requires: conflict detection (semantic similarity threshold), resolution protocol (voting, weighted scoring, supervisor arbitration), and deterministic tie-breaking (prefer higher confidence, more citations, or escalate to human)), and Observability Blind (agents operating as black boxes — "check the logs" is not observability. A 6-agent pipeline processes a request in 12 seconds. Which agent took 10 of those seconds? Which agent used 90% of the token budget? Which agent produced the error that was masked by a downstream retry? Without correlation IDs propagated across all agents, per-agent spans with latency/token metrics, per-agent error rates, and alerting thresholds — every production issue is a forensic investigation). Call once per agent architecture or pipeline before deployment