Skip to content
Canon

The Field · operating layer

Load a job. Follow a playbook. Refuse named failures.

The encyclopedia is the foundation. The Field is how agents use it — progressive disclosure at corpus scale. Do not ingest sixty articles. Identify the work, load that pack, copy the recipe, run the reviewer.

Load

Skills

34 SKILL.md packs. Index first, body on trigger.

Decide

Chooser

Four questions. A topology, a corpus, a refuse-list.

Audit

Reviewer

Paste a design. Score it against the five readiness questions.

Agents

llms.txt

Machine index. Then GET a job pack as markdown.

Skills
34
Jobs
11
Playbooks
15
Failures
22
Recipes
8
Corpus
60 / 56

Job packs

One pack per job. Load only that.

Coding agent

A harness that edits a repo: workspace, skills, plan/act, diffs, sandbox, verifier, human gate.

Load this pack, then the coding-agent playbook. Do not ingest the atlas. Start from a harness, not a blank graph.

Research agent

Long-horizon search with notes on disk, isolated researcher subagents, and a parent that synthesizes.

Parent synthesizes. Researchers write notes.md and return 30 lines. Never dump 40 hits into the parent window.

Support / ops agent

Router or handoff into specialists. Humans as tools. Durable state on the ticket. No group chat on prod.

Ticket is the blob. Handoff rewrites the brief. Refunds are ask_human. Do not start a crew.

Document / RAG agent

The hard problem is the corpus. Indexes, query engines, citations. Orchestration is secondary.

Prefer LlamaIndex (or equivalent) for the corpus. Retrieved text is untrusted data. Citations are part of the schema.

Browser agent

DOM as ACI, always sandboxed, last resort after an API. Computer-use is not the first tool.

API first. Browser last. Sandbox always. Action space is the product; keep it small and eval it.

Named workflow

The path is knowable. Chain, route, parallel, eval-opt. Code owns the edges. An agent island only at ambiguity.

Do not build an agent. Build a workflow. Put code gates between hops. Add a loop only where the eval fails without one.

Multi-agent team

A second agent is a context boundary. Supervisor, crew+flows, or handoff. Never unsupervised group chat on prod tools.

Roles must change tools or schema. Fresh windows. Structured returns. Bound the talk. Prefer supervisor or handoff over group chat.

MCP server

External capabilities as tools/resources/prompts. Pin, hash, review. Descriptions are an attack surface.

One job per tool. No secrets in descriptions. Clients pin this server. Do not instruct the model from a tool description.

Skill pack

Progressive-disclosure expertise. Write descriptions that trigger. Distill from traces. Do not flood the constitution.

Description first (what + when). Body as a playbook. references/ on demand. Trigger-test five phrasings.

Eval harness

Golden sets, rubrics, CI, red-team. The measurement plane the rest of the field sits on.

20–50 real tasks, structured asserts, CI on change, at least one injection/refusal case. Architecture waits on this.

Production agent product

Durable reducer, gateway from anywhere, traces, evals, HITL, owned prompts. The last 20%.

Name the five: termination, sandbox, context builder, eval set, pause API. One blob. One orchestrator. Gateway not a text box.

Wire format

Agents fetch /llms.txt, then a single job pack. Humans use the pages above. The JSON index is /api/canon.

GET /llms.txt
GET /api/canon/jobs/coding-agent?format=md
GET /api/canon/decide?path=unknown&work=coding&split=single&gate=sandbox
POST /api/canon/review  {"text": "…design…"}