Resilient Agent — Circuit Breaker & Error Recovery Workflow

The Problem

AI agents in production fail 13-41% of the time. Most fail silently — hallucinating answers, looping forever, or quietly dropping tasks.

The Solution

A prompt-layer resilience system. No code required — just instructions your agent follows when things go wrong.

What's Inside

• Error Recovery Protocol — 4-step decision tree: Classify → Act → Degrade/Escalate → Log
• Circuit Breaker Rules — Stop calling broken tools, auto-recover when they're back
• Error Log Template — Structured failure logging for debugging
• 4 error classifications: Transient (retry), Permanent (stop), Agent Error (re-approach), Unknown
• Cost circuit breaker — Prevents runaway spending from retry loops

Key Design Decisions

• 3x backoff (not 2x) — LLM APIs take 5-30s to recover
• Max 3 retries — diminishing returns after that
• "Re-approach" distinct from retry — agent calling tool wrong ≠ API being down
• Degrade preferred over escalate — "here's 80%" beats "I can't"

Files Included

• RESILIENCE.md — Core error recovery decision tree with classifications, retry logic, and degradation templates
• circuit_breaker.md — Per-tool and cost circuit breaker rules (CLOSED/OPEN/HALF-OPEN states)
• error_log_template.md — Structured failure logging format with weekly review questions
• examples/api_outage.md — Full walkthrough: handling a mid-task API outage
• examples/rate_limit_storm.md — Full walkthrough: surviving simultaneous rate limits across multiple tools

Add RESILIENCE.md and circuit_breaker.md to your agent's system prompt or workspace files. The error log template and examples are reference material. Works with any LLM that supports tool use — no code or infrastructure changes needed.