Why this exists

Netflix built Chaos Monkey because cloud apps that passed unit tests still went down in production — the failures were in the seams between systems, not the systems themselves. AI agents have the same problem, with a worse blast radius:

• A flaky tool returns malformed JSON → your agent hallucinates plausible-looking arguments and writes them to your database.
• A user pastes a "translate this" prompt that's actually IGNORE PREVIOUS INSTRUCTIONS → your support agent emails the customer your system prompt.
• A model upgrade changes how the agent retries a 429 → the agent enters an infinite loop and burns through your monthly token budget in 40 minutes.

These failures don't show up in unit tests because unit tests assume the seams work. agentfuzz deliberately breaks the seams.

What it does

Wrap your agent. Pick a fault profile. Run. Get a report.

from agentfuzz import Harness, faults

harness = Harness(my_agent)

harness.add(faults.ToolTimeout(rate=0.10))
harness.add(faults.MalformedToolResponse(rate=0.05))
harness.add(faults.PromptInjection.suite("owasp-llm01"))
harness.add(faults.CostSpiral(max_tokens=50_000))
harness.add(faults.LatencyJitter(p99_ms=8000))
harness.add(faults.PartialToolFailure())

report = harness.run(scenarios="tau-bench-airline", iterations=200)
report.html("./report.html")

You get:

• Pass-rate per fault category — "your agent survives malformed JSON 78% of the time but only 12% of timeout cases."
• Cost-blast radius — "fault X caused token usage to spike 14×."
• Tool-call failure modes — hallucinated arguments, retry storms, infinite loops.
• Prompt-injection survival — OWASP LLM01 suite results.
• Replay traces — the exact transcript that broke your agent, so you can fix it.

Install

pip install agentfuzz                       # core
pip install "agentfuzz[langgraph]"          # + LangGraph adapter
pip install "agentfuzz[crewai]"             # + CrewAI adapter
pip install "agentfuzz[autogen]"            # + AutoGen adapter
pip install "agentfuzz[all]"                # everything

60-second example

from agentfuzz import Harness, faults
from my_app import build_agent

harness = Harness(build_agent())
harness.add(faults.MalformedToolResponse(rate=0.2))
harness.add(faults.ToolTimeout(rate=0.1))

result = harness.run(iterations=50)
print(result.summary())
# >>> agentfuzz: 32/50 passed (64%)
# >>>   MalformedToolResponse: 8 failures
# >>>     - 5× hallucinated arguments
# >>>     - 3× silent corruption
# >>>   ToolTimeout: 10 failures
# >>>     - 7× retry storm (avg 14 retries)
# >>>     - 3× infinite loop killed at max_tokens

Fault library

More planned — see the roadmap.

Supported agent frameworks

• ✅ LangChain create_agent (1.x) — agentfuzz[langgraph]. The modern entry point. Wrap your tools with wrap_tools(), point LangGraphAdapter at the compiled graph.
• ✅ LangGraph create_react_agent (0.x) — same adapter; both APIs return a CompiledStateGraph we handle uniformly. See examples/langgraph_react_agent.py.
• ✅ CrewAI — agentfuzz[crewai]. wrap_tools() returns proxy crewai.tools.BaseTool instances; CrewAIAdapter(crew) drives the harness through crew.kickoff(). See examples/crewai_agent.py.
• ✅ AutoGen v0.4+ — agentfuzz[autogen]. wrap_tools() returns proxy autogen_core.tools.FunctionTool instances; AutoGenAdapter(agent) drives any agent / team exposing async run(task=...). See examples/autogen_agent.py.
• ✅ Plain Python callables — any Callable[[State], State]. Simplest way to try the tool.
• 🚧 PydanticAI, OpenAI Swarm, LlamaIndex — coming.

The adapter interface is small (is_available() + wrap()); PRs welcome.

Status

Alpha (v0.1). API will change. Built and tested on Python 3.10–3.13. The fault catalog is informed by production multi-agent deployments at enterprise scale — but every codebase fails in its own special way, so file issues when you find a fault we should ship.

Why I'm building this

I've spent the last decade architecting AI systems for enterprises — including multi-agent platforms running across 2,600+ production sites. The failures that hurt are almost never the ones the unit tests check for. They're the quiet, partial, half-degraded ones in the seams.

This is the tool I wish I'd had.

— Pavan Subhash Tirumalasetti

License

Apache 2.0. Use it commercially. Cite it in papers. Build a paid product on top. Just don't claim you wrote it.

Citing

If you use agentfuzz in research or production reports:

@software{agentfuzz,
  author  = {Tirumalasetti, Pavan Subhash},
  title   = {agentfuzz: Chaos engineering for AI agents},
  year    = {2026},
  url     = {https://github.com/SubhashPavan/agentfuzz},
}