Uploaded November 2026 | Updated November 2026, 1 day ago
Most LLM-powered recon tools look impressive in a 90-second demo and fall apart on real targets. They hallucinate, loop, go out of scope, double-fire the same endpoint, or die the moment a WAF or rate limit appears. This talk is about what it actually takes to run fifty-plus specialist agents in parallel for hours or days without the chaos.
We built a system where every agent is locked to the rails: a concrete tool, a strict pipeline phase, explicit prerequisites, timeouts, and sandboxes. The model does not drive. It rides. The binaries (and only the binaries) touch the target. This "Agents on Rails" approach eliminates freestyle LLM behavior while still letting the model do what it is good at—reasoning over evidence.
Most LLM-powered recon tools look impressive in a 90-second demo and fall apart on real targets. They hallucinate, loop, go out of scope, double-fire the same endpoint, or die the moment a WAF or rate limit appears. This talk is about what it actually takes to run fifty-plus specialist agents in parallel for hours or days without the chaos.
We built a system where every agent is locked to the rails: a concrete tool, a strict pipeline phase, explicit prerequisites, timeouts, and sandboxes. The model does not drive. It rides. The binaries (and only the binaries) touch the target. This "Agents on Rails" approach eliminates freestyle LLM behavior while still letting the model do what it is good at—reasoning over evidence.










