Running agents

Three reasons, and they are the only three.

• Context. Bulk reading belongs in a delegate, with a summary returned to the parent. A parent that reads a hundred files has spent its context on material it will not use again.
• Parallelism. Genuinely independent slices of work run at the same time, one delegate per slice, with the results aggregated.
• Isolation. Work that must not touch the parent's working tree gets its own.

Not a reason: hoping a second attempt on the same model will be better than the first. A re-attempt at identical configuration varies the sampling and nothing else, and where a genuine second perspective is needed the answer is deliberation or an independent checker rather than another agent.

Fan-out has two shapes and they fare differently here. Fan-out for coverage survives intact, because its value is parallelism and parallelism does not depend on diversity. Fan-out for competition, several attempts at one brief judged and grafted, depends on diversity and is therefore weakened rather than lost. Cross-family diversity is unavailable to this project. Diversity across models within the family is available, and arena supplies it (The pstack Collection).

The bound is worth stating wherever competition is used. Models trained together share more priors than models trained apart, so an intra-family bakeoff is a weaker instrument than a cross-family one and a stronger one than a single agent trying twice. Use coverage freely; reach for competition on design questions, and say which of the two you are relying on.

Parallel writers get their own worktree. This is not tidiness: two agents in one tree corrupt each other's work in ways that are hard to attribute afterwards, and the failure surfaces long after the cause.

Each environment carries its own worktree, its own ports, its own .env, and its own database, provisioned through compass rather than by hand. compass fleet is the supervisor's view: what every worktree is doing, which branch, which task, which pull request.

Two shared resources cut across the isolation and must be handled explicitly, because a worktree does not protect them: the git stash stack, which is shared across every worktree in the repository, and any remote the agents push to. Neither is a place for a delegate to improvise.

Five fields, and the model is deliberately not among them.

The model is inherited from the session that spawned the agent. Roles are not routed to model tiers by default: the operator chooses one model and every delegate uses it, which keeps the configuration small and the behaviour predictable. Where an instrument depends on diversity, as arena and interrogate do, the model is named explicitly for that call and the bound on what it buys is stated.

An agent left running needs four things the attended case can do without.

The parts of this architecture that matter most to an unattended run are the ones that bind rather than advise: the finish condition, the authority boundary, the read-only Systems, the isolation, and the independent checkers. Everything advisory, including the regulatory functions themselves and the default visibility at most levels, helps an agent choose well while someone is watching and does nothing at three in the morning.

The fleet is live state rather than a table. compass fleet reports what every worktree is currently doing: its branch, its task, its open pull request and its synchronisation with the trunk. That report is the inventory, and a written copy of it would be wrong within the hour.