The agents are answering. The code is committed. The watchdog is green. So why has the project stopped moving?
Agent orchestration resembles a bee colony: success depends on the coordination of the whole. In an agent system, individual workers can remain responsive and apparently healthy while the coordination needed to carry the work forward has collapsed. I call this operational pattern Context Collapse Disorder (CCD), deliberately echoing Colony Collapse Disorder. The analogy is organizational, not biological.
Drawing on a reconstructed multi-agent development incident, I follow completed work that a relay could see but the coordinator had not accepted, and local tests that passed while a project-level integration requirement remained unmet. The distinction matters for defenders: a healthy component does not certify a healthy workflow.
The security risk can outlast the incident. An unsupported conclusion can survive in a polished report, memory, or handoff, then return as trusted input to another agent. I call this risk semantic reinfection. A hash can prove that a document has not changed; it cannot prove that its conclusions were correct.
Using the incident record and a synthetic Bee Colony test harness, I examine missing requirements, failed handoffs, and unsupported completion claims separately. The experiments include both successful and failed attempts to recreate potential explanations. Just like the bee version of CCD, multiple factors amplify each other rather than a single cause driving the outcome.
Attendees will leave with practical checks for tracing a task from assignment to acknowledgment and acceptance, validating inherited claims against primary evidence, and preserving a recovery trail. When an agent says “done,” ask: done where, received by whom, and accepted against which requirement?
About the Presenter: Sebastien Doucet
Sebastien Doucet, aka TiGa, is a security researcher, reverse-engineer and instructor with over 20 years of experience, including training for law enforcement, military and government audiences. He is the founder of Brewster Jennings Canada and has presented at REcon, SecTor and other conferences internationally. His current work focuses on agent reliability, model post-training and experimental systems that learn, retain skills and train other models. Yes, that's a thing. He brings an instructor's approach to AI research: test what was learned, check what survives the next lesson and keep the evidence when an experiment disagrees with the theory.
