Skip to main content

Operating loop

Use one mental model everywhere in the docs: Farmslot is the loop from intent to approved evidence, then back into memory.

Stages

StagePurpose
Roadmap + memoryCapture raw ideas, external-source signals, product intent, constraints, and previous run learnings in one place.
Backlog refinementWork with the operator to clarify, prioritize, break down, and scope vague intent into dispatchable work.
Dispatch queuePick ready work when slots and runners are available.
Isolated slotGive the agent a clean project workspace, resources, and health checks.
Agent runnerRender the project-owned prompt and let Claude, Codex, Cursor, browser-automation runners, or custom runners implement.
Observe + steerStream terminal/run state and allow typed or voice intervention.
Recipe validationProve behavior with deterministic checks and artifacts.
Cross-runner reviewUse independent reviewers to catch different classes of issues.
Human approvalKeep the operator responsible for judgment and final acceptance.
Replay + improvementConvert useful outcomes into better prompts, templates, recipes, and evals.

Adoption layers

The loop can be adopted incrementally:

  1. start with one review bottleneck and the proof artifact that would make it easier to approve;
  2. encode that proof as a narrow recipe/evidence package;
  3. add isolated slots for safer parallel work;
  4. route work through the gateway and Command Center;
  5. 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.