— 01
Roles, not tasks.
Define your org in YAML. Each role has a persona, a reporting line, a toolset, and persistent memory. Non-engineers can read and amend it.
Conclave brings roles, deliberation, and auditability to multi-agent systems built on Anthropic primitives. A thin, opinionated framework that treats organizational coordination — not a graph — as the first-class object.
Most multi-agent frameworks model tasks or graphs. Real organizations don't work that way. They work through roles that hold persistent accountability, deliberate together, and leave an audit trail. Conclave makes that the primitive.
Define your org in YAML. Each role has a persona, a reporting line, a toolset, and persistent memory. Non-engineers can read and amend it.
A classifier decides per task: repetitive work runs a Haiku self-correction loop, novel work goes to Sonnet. Savings on production workloads run 60–80% versus frontier-only.
Every inter-agent message lands in a JSONL trail with sender, recipient, reasoning, and cost. Render it as a Mermaid sequenceDiagram with conclave trail view — postmortem-ready in one command.
A single-file dashboard ships with the framework — four MetricCards, four ChartCards, a live SSE activity stream, and a dual feed. Visually grounded in the Anthropic design system.
A Conclave is a small group of Claude-backed agents, each anchored to a role, working through a structured deliberation until a concrete output is produced. The configuration below is the entire specification.
org:
name: "Product Squad"
deliberation: hierarchy
backend: managed_agents # anthropic | managed_agents
agents:
- role: CPO
persona: "Strategic. Defines scope. Validates value."
tools: [notion, slack]
force_model: claude-sonnet-4-6
- role: TechLead
persona: "Tight specs. Two options, not one."
reports_to: CPO
tools: [github, linear]
- role: QA_Engineer
persona: "Adversarial. No approval without test plan."
reports_to: TechLead
tools: [github, browserbase]
Running it is one command. Add --dry-run
to simulate the pipeline end-to-end — trail, cost accounting,
deliberation — without spending a single token.
$ pip install conclave-agents
$ conclave init --template product-squad
$ conclave run "Spec the checkout API" --dry-run
◆ Conclave · Deliberation started
[user] → CPO
Spec the checkout API
[CPO] → TechLead
Please pick this up.
↳ delegating to TechLead
[TechLead] → QA_Engineer
Your turn. Flag blockers early.
↳ delegating to QA_Engineer
✓ Output from QA_Engineer
◆ Deliberation complete
Trail → .conclave/trail_20260419.jsonl
haiku 865 in 359 out $0.0021
sonnet 988 in 328 out $0.0079
─────────────────────────────────────
TOTAL $0.0100
BASELINE (all-Sonnet) $0.0159
SAVED $0.0059 (36.9%)Conclave does not reinvent the agent. It wraps Managed Agents sessions in a coordination layer, binds tool access to MCP servers, and routes across the Claude model family. Every Anthropic release is a step forward, not a rewrite.
Figure 1 — YAML configuration is the only input. The runtime coordinates agents, logs every decision, and calls Anthropic primitives through a swappable backend.
A live dashboard ships with the framework — single file, dark surface, built on
the tokens npx getdesign add claude produces. It gives you the state
of the deliberating org in one screen: metric cards, 14-day charts, and a dual
feed that reacts to the SSE stream off the Decision Trail.
$ conclave dashboard
◆ Conclave Dashboard → http://localhost:7777
The dashboard is powered by the same JSONL trail that conclave trail view
renders as Mermaid. Pollers that target Anthropic agents typically refetch every
few seconds — Conclave emits events on /api/events as soon as the
trail file changes, so a new handoff appears in under 500 ms.
Warm dark palette (#141413 / #30302e), terracotta accent
(#c96442) reserved for live indicators and CTAs, serif headlines
(Fraunces as Anthropic Serif proxy). No cool grays — every neutral has a
yellow-brown undertone, consistent with the Claude design system.
A persistent agent's history grows with every turn. We refuse to pay a Sonnet bill to format a table. The router runs a Haiku classifier on each task, then dispatches — Haiku loop for the routine, Sonnet for the novel, Opus only when the domain warrants it.
Figure 2 — A task is classified on Haiku, then dispatched. Low-novelty work runs a Haiku self-correction loop; novel work goes to Sonnet; Opus is reserved for explicit, high-stakes domains.
Every handoff is a message with explicit sender, recipient, type
(delegation, handoff,
escalation, output)
and reasoning. The diagram below is a real deliberation rendered from
the Decision Trail.
Figure 3 — Four messages across three roles. Each line is one entry in the JSONL Decision Trail, timestamped, with the model that produced it.
Multi-agent frameworks have proliferated. Most rebuild primitives Anthropic already ships. Conclave chooses the opposite: stay thin, lean on Managed Agents and MCP, and add the one thing nobody else does — the organizational layer.
| LangGraph | CrewAI | AutoGen | Conclave | |
|---|---|---|---|---|
| Agent primitive | Graph node | Task worker | Chat participant | Organizational role |
| Coordination | Explicit DAG | Linear pipeline | Free-form chat | Deliberation mode |
| Memory | Checkpointer | Per-task | Per-conversation | Persistent, per-role |
| Cost control | Manual | Manual | Manual | Router: Haiku / Sonnet |
| Audit | Optional traces | Logs | Chat history | Decision Trail, always on |
| Config surface | Python graph | Python + YAML | Python | YAML-first |
An org isn't a DAG — it's a living system of accountabilities. — Conclave README · Philosophy
Conclave's backend abstraction means switching from the stateless
messages.create() path to native
Managed Agents sessions is a one-line YAML change. When the beta
goes GA, the migration is a no-op.
org:
backend: managed_agentsInstall, initialize from a template, simulate the full pipeline — then swap in an API key when you're ready for a real run.
$ pip install conclave-agents
$ conclave init --template product-squad
$ conclave run "Your goal" --dry-run$ conclave dashboard
◆ Conclave Dashboard → http://localhost:7777$ conclave trail view --latest > trail.md
# Mermaid sequenceDiagram, renders on GitHub$ conclave benchmark
Category Haiku Sonnet Conclave Quality
──────────────────────────────────────────────────
repetitive $0.008 $0.062 $0.009 97%
operational $0.011 $0.081 $0.024 91%
technical $0.014 $0.098 $0.071 96%
strategic $0.009 $0.071 $0.065 98%
──────────────────────────────────────────────────
TOTAL $0.042 $0.312 $0.091 94%~70% saved versus all-Sonnet at 94% quality parity. Every task lives in the repo — reproduce with one command.