Failure atlas
The operating system was written by the failures it could not afford to repeat.
The most durable orchestration knowledge comes from measuring a failure, naming its cost, and building the smallest enforcement that prevents a second occurrence.
Read the original on XThe evidence card
This is the central evidence archive behind Citadel: 27 documented failures across agents, waves, worktrees, hooks, interfaces, and performance. It turns an abstract promise about agent reliability into a catalog of specific breaks and the infrastructure each one produced.
The model
The visual is an editorial compression of the source essay: enough structure to orient a reader, without pretending that the model replaces the underlying evidence.
The argument
The site version keeps the source's operational insight while making its sequence easier to inspect, share, and connect to the rest of the archive.
The atlas begins with confident wrong work, wiped context, truncated scope, agent collisions, lost waves, replacement fleets, and design loops that never terminate. These are coordination failures, not syntax failures.
Instruction bloat, animation loops, backdrop blur, focus-stealing hooks, and baseline type errors compound across individually reasonable edits. The fix is a system-level signal, not another paragraph in a prompt.
Invisible features, state feedback loops, routing drift, hardcoded IDs, and template-shaped design all pass shallow checks. The failures expose where static verification ends and lived use begins.
The boundary
The scars are the system. Everything else is commentary.
The 27 cases are a local evidence set from one large codebase. They are transferable patterns, not a population-level claim about every agent system.