Skip to content

Work / canon / Wiki / Decisions

agent-pipeline

Decisionverified 2026-11-02

Agent pipeline

Choice: Canon also runs work, not only measures it. Refactor canon and canon-cli in place into one configurable pipeline: Problem → Idea → POC (optional) → Plan → Feature/Bug → Design (optional) → Development → Review → Release. This is only the default: a team can add, remove or change any stage, record kind or connector, at any time, in config. A trigger, or a person, opens a Problem.

  • Split: what runs on one machine from the repo goes in canon-cli, and stays generic. What needs a server or other people goes in canon. See editions.
  • canon-cli: pipeline format, local runner, manual, scheduled and polling triggers, doc types, checks, read-only connectors with the user's own keys, a local cost log.
  • canon: webhooks, hosted runs, duplicate removal across repos, browser approvals, connectors for a customer's own third-party tools with stored logins, audit log, dashboards.
  • Teams: organization, products, teams and spaces follow product-model.
  • Babul stays the design tool; Canon calls it for design deltas and Verify. Neither copies the other's code.
  • Safety: an agent never approves. Text from a trigger is data, never instructions, and the agent reading it has no write access. Connectors start read-only. Each trigger has a spend cap.
  • Order: the owner's own products are the first customer. Outside sales come after two weeks of real use on one live product. No paid-site work until T040 and T043 are done. Why: the owner, 2026-11-02: products should move while they are away, and they only approve the gates. Shipping is the bottleneck and nothing earns yet. Stages, gates, search and weigh already exist; only triggers, duplicate removal and a runner are missing (flow measures but does not run). Rejected: a new repo (rebuilds what works, and leaves a parallel copy) · an n8n-style canvas now (n8n already runs steps with 500+ integrations; Canon's edge is the reason, proof, approval and result) · built-in marketing stages now (AskCal's decision waits for its review around 2026-11-28) · building on n8n as the product (its licence needs a commercial deal when a product's value comes mainly from n8n). Revisit: the canvas, once a client asks to edit a pipeline themselves. Built-in marketing stages, after the 2026-11-28 review. Continuing at all on 2027-01-31: it must have saved measurable hours and one outside company must have paid; if not, stop adding to it and record why.