Skip to content

How It Works

AgentBreak is a transparent proxy. It sits between your agent and the services it depends on, intercepting traffic and injecting controlled faults.

Agent  →  AgentBreak (localhost:5005)  →  Upstream LLM / MCP
               ↑
         scenarios.yaml (fault rules)

The agent talks to AgentBreak as if it were the real service. AgentBreak either forwards to the real upstream (proxy mode) or returns synthetic responses (mock mode), optionally corrupting, delaying, or replacing responses along the way.

Supported API formats

AgentBreak exposes three endpoints, all backed by the same fault injection engine:

Endpoint Format
POST /v1/chat/completions OpenAI
POST /v1/messages Anthropic Messages
POST /mcp MCP (JSON-RPC over HTTP)

What gets tested

AgentBreak focuses on infrastructure-level failures, not semantic attacks. The goal is to answer: "Does my agent handle real-world service degradation gracefully?"

Category What it simulates Example
Availability Service outages, rate limits http_error with status 503 or 429
Latency Slow responses, network delays latency with min/max milliseconds
Timeouts Requests that never complete timeout with min/max milliseconds
Corruption Broken payloads invalid_json, empty_response, schema_violation
Content drift Unexpected response content wrong_content, large_response

Behavioral detection

Beyond injected faults, AgentBreak passively monitors:

  • Duplicate requests — same payload sent more than once (fingerprint-based)
  • Suspected loops — same payload sent 3+ times, indicating the agent is stuck
  • Upstream failures — real errors from the upstream service (not injected)

Scoring

After a run, the scorecard gives you:

  • Resilience score (0-100) — starts at 100, deductions for faults, failures, duplicates, and loops
  • Run outcome — PASS (clean), DEGRADED (partial failures), or FAIL (all failed or looping)
Deduction Amount
Per injected fault -3
Per upstream failure -12
Per duplicate request -2
Per suspected loop -10