Browser Chaos Engineering: Direct CDP Fault Injection & Race Telemetry
Published: 2026-10-30 · Engineering
Synthetic agent benchmarks assume clean networks and cooperative DOM trees. We built wt-choas to shatter that assumption: injecting socket timeouts, 4x CPU slowdowns, multi-click race bursts, and boundary fuzzing over raw CDP — forcing agent self-healing loops to prove their resilience before production.
## The Genesis: Why Sunny-Day Agent Testing Fails
Almost every browser agent framework looks miraculous in a staged demonstration. You point the model at an e-commerce checkout or a dashboard form, it fires three clean tool calls, and the task completes in twelve seconds.
Then you run that exact agent pipeline against a live production web application under real-world conditions, and everything collapses:
1. A payment gateway request stalls for 1,800ms over a jittery cellular connection; the agent times out and hallucinates that the transaction failed, submitting a duplicate charge.
2. A debounced search dropdown renders asynchronous suggestion elements with a 120ms animation delay; the agent clicks the stale coordinates and clicks through to a 404 page.
3. An unhandled \