Analysis updated 2026-05-18
Test whether an app handles a hung or timed-out LLM call gracefully
Simulate a rate-limit storm to check retry and backoff logic
Catch silent output corruption that normal monitoring would miss
Verify a multi-agent system has safeguards against infinite loops or tool failures
| deepagentlabs/agentic-chaos | 1lystore/awaek | 47cid/wp2shell-lab | |
|---|---|---|---|
| Stars | 13 | 13 | 13 |
| Language | Python | Python | Python |
| Setup difficulty | easy | moderate | moderate |
| Complexity | 3/5 | 2/5 | 4/5 |
| Audience | developer | vibe coder | researcher |
Figures from each repo's GitHub metadata at analysis time.
Installs standalone via pip install agentic-chaos, with no required dependency on other packages.
agentic-chaos is a Python testing toolkit that deliberately breaks an application's calls to language models and multi-agent systems, then reports exactly what happened, so a developer can see how their code handles failure before it happens in production. It installs on its own with pip and does not require any other package to work, including a related project called AgenticLens, though it can optionally feed its results into that tool if someone uses both. At the center of the library is a function called chaos_call, which wraps a normal function call and only misbehaves when it is running inside an active chaos session. Outside of that session, it passes straight through as if the library were not there, so the same instrumented code stays safe to ship. The three faults available in this first version are a token timeout that hangs a call before raising an error, a rate limit storm that fails the first few calls before letting things through normally, and silent degradation, which lets a call succeed but quietly corrupts its returned text while keeping the timing and token counts looking normal. That last one is called out as the hardest to catch, since nothing in ordinary cost or latency monitoring would look wrong. A second version adds agent level faults for testing systems made of multiple cooperating agents: forcing a specific tool call to error, time out, or return empty, corrupting shared memory between agents by truncating or garbling it, and forcing an agent to loop past its intended turn limit to check whether it has a safeguard against that. A wrap_tool helper applies these faults transparently to existing tool functions while also recording which agent called which tool, building a map of how the system is actually wired together. Everything can be run from the command line against a real script, printing a table of what faults fired and what happened, with the option to save a standalone JSON report that needs no other tool to read. A planned third version, a prompt and model drift detector, is listed on the roadmap but not yet built.
A Python toolkit that deliberately injects failures into LLM calls and multi-agent systems to test how well an app handles them.
Mainly Python. The stack also includes Python.
Setup difficulty is rated easy, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.