What this is
I made this repo as the first small demo in a broader attempt to build and share things with agents. The point is the orchestration itself, not a product that someone would install and use. The central experiment is a three-layer shape: a top-level director creates tasks, those tasks start worker agents when no agent is assigned, and the workers produce separate contributions toward one shared goal.
How the experiment works
The configuration lives under .agents/. The hub is the coordination point, with task creation, agent sessions, isolated worktrees, and a shared artifact area. The run documented here used a director called Atlas to distribute a brief for six landing-page variants. Each worker had a separate worktree, while finished files and JSON manifests were published into a common artifact directory. That separation turned out to be important: work in progress stayed isolated, and the shared directory acted as a blackboard for composing independent outputs without git merges or file collisions.
The run produced six parallel workers in roughly a second, then retried two variants on another model after the first provider stalled. Two of six variants eventually shipped as fully offline HTML pages. The surrounding notes and docs/ files keep the useful evidence: transcripts and hub state showed that the fan-out itself held up, workers self-validated their pages, and the director was able to assemble a partial result.
Current state
This is still an experiment and a fairly honest one. The run exposed several unfinished parts of the orchestration layer. Empty, zero-token provider turns were recorded as successful completions, so failed workers looked healthy. There was no reliable liveness signal to distinguish a slow agent from a dead one, and tasks could enter in_review without a path to a terminal state. The director also briefly proposed doing the workers' work itself, which made the need for explicit "wait, retry, or ship partial" constraints clear.
What remains is mostly infrastructure around the core idea: failure-aware turn handling, task completion and approval transitions, last-token or heartbeat reporting, resolved spawn configuration in the event history, change notifications for shared artifacts, and validation of the output contract. For now, the repo is a working record of the first orchestration test and the design questions it surfaced.