What is Papercup?

Papercup is the operating system of an autonomous company. It has two layers:

• Org layer: five director agents — Business, R&D, Technology, Management, Marketing — coordinating via typed inbox/outbox messages. They route work, allocate budget, kick off projects, and surface market signals. They don’t write code.
• Coding harness: each project (e.g. music-app, sales-dashboard) runs its own bash-orchestrated agent loop with planner / worker / validator / etc. This is the layer that physically ships features. The harness predates the org layer and runs on its own.

The directors decide which projects exist and what they should do. The coding harnesses inside each project execute the actual coding work. You write directives; humans observe and intervene at named checkpoints when either layer asks for a decision.

These docs cover both. The Org layer section covers the director coordination layer. The Architecture + Decisions sections cover the coding harness internals.

Concretely, what is “the harness”?

A bash loop (run.sh) that orchestrates a small fleet of role agents. Each iteration:

1. Asks the orchestrator what to do next (NEXT_WORKER F-023, NEXT_VALIDATOR F-022, DONE, ESCALATE …, CHECKPOINT …).
2. Invokes that role with a fresh context. The prompt is a stable per-role markdown file plus auto-injected memory layers (cross-mission identity + per-mission summary).
3. Captures stdout/stderr, parses the decision verb, mutates .harness/features.json, and snapshots state.
4. Loops.

13 roles ship in the box: planner, plan-reviewer, worker, validator, crosscheck, ui-qa, debugger, architect, curator, documenter, product, supervisor, orchestrator. They all have specialised prompts; none of them carry context between invocations beyond what’s on disk.

What makes Papercup different from “AI agents that code”

• It runs unsupervised. No human in the inner loop. Iterations have no upper bound except the cost cap (which is currently disabled).
• It admits when it’s stuck. Named checkpoints (Hermes-inspired) and validator escalations halt the run; humans grant via UI.
• It writes its own scope. A product role periodically diffs SPEC.md against GOAL.md and proposes new features. Operators triage the proposals.
• It learns. The curator maintains a per-role identity file (~/autonomous-harness/identity/<role>.md) that survives across missions.
• It races against itself. Optional competition mode (Conductor-inspired) spawns N workers on the same feature; a validator picks a winner.