Skip to content

Orchestrator + Specialists

One coordinator picks which specialist handles each sub-task. The specialists never talk to each other — only to the orchestrator. Think project manager + team.

Orchestrator pattern — Coordinator at top, three Specialists below, parallel dispatch with results merging back

What it is

The coordinator uses its model to pick which specialists handle a task, then runs each selected specialist's full agent loop and correlates the results. When the coordinator dispatches to multiple specialists in one turn they run in parallel (bounded by max_parallel_specialists).

Each Specialist is its own self-contained agent. Its fields:

  • a name — what the coordinator calls it by
  • a specialist_type — a short type tag (e.g. "research")
  • a description — what the specialist is good at (the coordinator reads this)
  • a system_prompt — the specialist's own instructions
  • its own tools and model
  • an optional confidence_threshold (default 0.85) — the bar the specialist's self-estimated confidence is measured against

When to use it

  • ✅ The work splits cleanly into expert domains (Research, Data, Writing; or Triage, Forensics, Containment).
  • ✅ You want one place to attribute decisions to — the coordinator.
  • ✅ Specialists need their own playbooks, skills, or models (a cheap model for triage, a strong one for compliance).
  • ✅ Auditability matters — the dispatch log tells you who ran and why; for consequential actions, route the tool through admit() so every decision lands on the hash-chained AuditTrail.

When NOT to use it

  • ❌ The flow is linear, not delegated — use Composition.
  • ❌ No central coordinator should exist; agents should self-organise — use Swarm.
  • ❌ The conversation itself moves between roles — use Handoff.

Code

from tulip.multiagent import Specialist, create_orchestrator

model = "anthropic:claude-sonnet-4-6"

research = Specialist(
    name="research",
    specialist_type="research",
    description="Searches sources. Collects evidence. Summarises what it found.",
    system_prompt="You are the Research specialist.",
    tools=[web_search, fetch_page],
)

data = Specialist(
    name="data",
    specialist_type="data",
    description="Queries the warehouse. Reconstructs timelines. Quantifies impact.",
    system_prompt="You are the Data specialist.",
    tools=[query_metrics, query_warehouse],
)

writing = Specialist(
    name="writing",
    specialist_type="writing",
    description="Drafts and publishes the report. Only after research + data agree.",
    system_prompt="You are the Writing specialist.",
    tools=[draft_report, publish_report],  # ← idempotent writes
)

orchestrator = create_orchestrator(
    name="coordinator",
    specialists=[research, data, writing],
    model=model,                           # the coordinator's routing model
)
orchestrator.system_prompt = (
    "You are the coordinator. Delegate source-gathering to research, "
    "metrics to data, and only after both report back call writing."
)

result = await orchestrator.execute(
    "Why did checkout conversion drop last week? Investigate and draft a report.",
)

create_orchestrator registers the specialists and propagates the coordinator's model into any specialist that doesn't carry its own. execute() is async — await it (or wrap in asyncio.run). The same shape runs an incident commander: triage, forensics, and containment specialists, with containment gated behind the other two.

What runs in parallel

Specialists fire concurrently when the coordinator dispatches to several of them in one turn. So when the coordinator says "in parallel: research, pull the release notes for deploy-77; data, pull last week's funnel metrics" — both specialists run at the same time and their results merge back before the coordinator's next Think.

Confidence thresholds

Each specialist carries a confidence_threshold (default 0.85). Every SpecialistResult reports a self-estimated confidence, so you can compare it against the threshold and decide whether to trust the output or route the sub-task to another expert:

Specialist(
    name="data-quality",
    specialist_type="data_quality",
    description="Audits metric definitions and flags unreliable numbers.",
    system_prompt="You audit metrics before they are cited.",
    tools=[query_warehouse, profile_table],
    confidence_threshold=0.7,    # the bar this specialist's
                                 # self-estimated confidence is held to
)

Notebooks

Source

multiagent/orchestrator.py — Orchestrator, Specialist.

See also