Skip to content

Docs home · Usage · Councils · Cost guide · Orchestration · Architecture eval

Orchestration Backends

The council never requires a single runtime. At session start, a capability check resolves a backend per mode and degrades gracefully — it never hard-exits.

The three backends

Backend What it is Needs
workflow Deliberation/audit runs as a background dynamic workflow — a script orchestrating subagents, with round texts kept on disk and out of your session's context Claude Code ≥ 2.1.154, workflows enabled (default on Max/Team/Enterprise; /config toggle on Pro)
teams Each council member is a live agent-teams teammate you can message directly CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 (experimental)
sequential The conductor drives the same rounds with plain agent calls nothing

All three produce identical on-disk artifacts. Sequential is slower and chattier; it exists so the plugin works everywhere.

Which backend runs what

Mode / phase Preferred Degrades to Why
brainstorm, quick inline agent calls 3 one-shots; orchestration runtime is overhead
standard / auto / deep / guided deliberation workflow teams → sequential The 3-round structure is deterministic — script-shaped. Background execution, in-session resume, hard token budget
--meet cross-examination teams workflow-style fixed phases → sequential Direct agent-to-agent Q&A and live steering are what teams are for
Path A team execution teams sequential over the task dependency list Shared task list, dependency unblocking, per-task plan approval
Deep audit (5D) / --audit workflow conductor-driven waves Loop-until-convergence with a budget ceiling; replaces checkpoint/respawn machinery

The resolved backend is printed when it differs from the preferred one, and recorded in session.md so --resume continues on the same path.

Why deliberation moved to workflows (v1.2)

Anthropic's framing for choosing between primitives is who holds the plan: with subagents and teams, Claude decides turn by turn; with a workflow, the script decides. The council's deliberation — survey the codebase once for everyone, positions in parallel, pick tension pairs, paired rebuttals, converge, synthesize — is a fixed structure. Encoding it as a script bought:

  • No experimental flag for the core flow — workflows are available on all paid plans.
  • Context hygiene — a 7-agent, 3-round deliberation's full text never enters your session; only the structured Decision Log and Tension Resolutions come back.
  • Background execution — your session stays responsive while rounds run; watch progress via /workflows.
  • Resumability — completed agents return cached results if a run is interrupted.
  • A hard token budget — the mode's budget (e.g. ~750K standard, ~2M audit) is a ceiling, not a suggestion.

The scripts ship in the plugin at references/workflows/council-deliberation.template.js and council-audit.template.js. The conductor writes a per-session copy with one marked line rewritten to a literal INPUT object (roster, idea, skill content, model routing) and runs that copy via the Workflow tool's scriptPath; the logic below the INPUT line is byte-for-byte identical, so the same scripts serve all four councils. Injecting literals replaced an earlier attempt to pass everything through the runtime args global, which proved unreliable for large nested payloads — a failed run now degrades to teams/sequential rather than improvising a substitute script.

What teams still do better — and why --meet and Path A prefer them: teammates are persistent (they remember their own Round 1 reasoning), they message each other directly, and you can jump into any teammate's session mid-flight (Shift+Down). Workflow agents compensate by re-reading their own prior round files from disk ("it is YOUR prior work"), which works because every round has always been persisted.

Degradation matrix

Mode WF ✓ Teams ✓ WF ✓ Teams ✗ WF ✗ Teams ✓ Neither
brainstorm / quick inline inline inline inline
standard / auto / deep / guided workflow workflow teams sequential
meeting teams workflow phases, no live steering teams sequential
Path A execution teams sequential teams sequential
audit / 5D workflow workflow conductor waves conductor waves

Troubleshooting

Orchestration: sequential notice on every session — neither runtime is available. Update Claude Code to ≥ 2.1.154; check Dynamic workflows isn't off in /config (or disabled by CLAUDE_CODE_DISABLE_WORKFLOWS / org managed settings); or export the teams flag. Sequential still works.

--meet says live steering is unavailable — that specific feature needs CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 exported in the shell that launched Claude Code (then restart).

Deliberation workflow fails to launch — the engine treats the failure as detection and re-runs on teams (if enabled) or sequential. If it keeps happening on a plan that should support workflows, check claude --version and /config.

Teams quirks (only relevant when the flag is on): /resume doesn't restore in-process teammates, shutdown can be slow, one team at a time. These now affect only --meet steering and Path A — not the core flow.


Next: Architecture eval — the design record behind these choices.