Agents
RDF uses seven universal agents, plus two variants that differ only in effort. Their behavior comes from each project’s governance files, not from domain knowledge baked into prompts. A QA agent reviewing a bash firewall and one reviewing a Python pipeline follow different rules because their governance differs.
Roster
| Agent | Model · effort | Role | Access |
|---|---|---|---|
rdf-planner | fable · high | Design specs and implementation plans. /r-spec and /r-plan run this protocol inline; direct dispatch also works | Read and write |
rdf-dispatcher | opus · high | Orchestrate plan phases and TDD cycles | Read and write |
rdf-engineer | opus · xhigh | Implement through the governance protocol | Read and write |
rdf-engineer-focused | opus · medium | Variant for scope:docs and scope:focused phases. Failures escalate to rdf-engineer | Read and write |
rdf-qa | opus · medium | Verification gate | Read and execute |
rdf-reviewer | opus · xhigh | Adversarial sentinel review | Read and execute |
rdf-reviewer-challenge | opus · high | Variant for pre-implementation challenge review | Read and execute |
rdf-uat | opus · medium | User acceptance from an end-user persona | Read and execute |
rdf-claims | opus · high | Fresh-context check of text before it is posted: PR bodies, commit messages, changelogs, issue and chat drafts (direct dispatch) | Read and execute; Edit disallowed, and its prompt limits writes to its report, evidence and scratch |
Why agents stay narrow
Agents are context buffers, not general-purpose helpers. A QA agent loaded with only the diff, the test results and the verification protocol catches issues that an agent carrying the whole session would miss.
Routing
Routing lives in adapters/claude-code/agent-meta.json, which sets model, effort and variants. Variants generate from the base agent’s canonical body. The Agent tool cannot pass effort per call, so each effort level is its own agent.
| To change | Do this |
|---|---|
| Effort for one session | Use /effort, which saves a level per model |
| Effort for every agent | Set CLAUDE_CODE_EFFORT_LEVEL |
| The planner’s model | Set its "model" in agent-meta.json and regenerate |
Spec and plan sessions are tuned for claude --model fable --effort high, and builds for claude --model opus. No Fable access? Set the planner’s "model" to "opus" and run rdf generate claude-code.
/r-util-claudemd-review deliberately fans out to Sonnet for bulk transcript reading.
More detail
WORKFORCE.md has the organization chart, per-agent notes and common workflows.