Operating loop
Use one mental model everywhere in the docs: Farmslot is the loop from intent to approved evidence, then back into memory.
Stages
| Stage | Purpose |
|---|---|
| Roadmap + memory | Capture raw ideas, external-source signals, product intent, constraints, and previous run learnings in one place. |
| Backlog refinement | Work with the operator to clarify, prioritize, break down, and scope vague intent into dispatchable work. |
| Dispatch queue | Pick ready work when slots and runners are available. |
| Isolated slot | Give the agent a clean project workspace, resources, and health checks. |
| Agent runner | Render the project-owned prompt and let Claude, Codex, Cursor, browser-automation runners, or custom runners implement. |
| Observe + steer | Stream terminal/run state and allow typed or voice intervention. |
| Recipe validation | Prove behavior with deterministic checks and artifacts. |
| Cross-runner review | Use independent reviewers to catch different classes of issues. |
| Human approval | Keep the operator responsible for judgment and final acceptance. |
| Replay + improvement | Convert useful outcomes into better prompts, templates, recipes, and evals. |
Adoption layers
The loop can be adopted incrementally:
- start with one review bottleneck and the proof artifact that would make it easier to approve;
- encode that proof as a narrow recipe/evidence package;
- add isolated slots for safer parallel work;
- route work through the gateway and Command Center;
- add backlog/dispatch policy, cross-runner review, and replay/eval loops.
Visual validation, executable recipes, and proof artifacts are the trust layer. The gateway is the control plane. The full loop is the product.