Orchestrator (Agent Teams)
Run teams of AI agents that collaborate on a goal using your Crow data — searching memories, reading projects, writing summaries — and schedule recurring jobs that keep your knowledge base tidy while you sleep.
What is this?
The orchestrator turns a single request into a coordinated team effort. You give it a goal in plain language; a coordinator agent breaks it into tasks and hands them to worker agents, each equipped with just the Crow tools its role needs. The team shares findings and the coordinator synthesizes a final answer.
Why would I want this?
- Multi-step research — one agent digs through your memories and project sources while another writes up what it finds
- Recurring maintenance — a nightly pipeline finds duplicate or conflicting memories so your recall stays sharp
- Scheduled digests — a daily summary of activity, or a weekly review of all active projects, delivered as a memory without you asking
Running a Team
Ask your AI client to orchestrate — it uses the crow_orchestrate tool behind the scenes:
"Orchestrate a research team on: what do my notes say about FERPA compliance?"The run starts in the background and returns a job ID. Ask for status anytime:
"Check on that orchestration"
→ crow_orchestrate_status({ jobId: "..." })Choosing a Team Preset
Presets are ready-made team configurations. Ask "list the orchestrator presets" (crow_list_presets) to see everything available on your instance. The general-purpose ones:
| Preset | What it does |
|---|---|
research | One agent searches memories and projects, another synthesizes the findings |
memory_ops | A single analyst searches, consolidates, and organizes memories |
full | Researcher + memory writer + synthesizer with broad tool access |
code_team, vision_team, deep_synthesis | Specialized teams for code, image, and deep-synthesis work |
Additional presets power Bot Builder bots and instance-specific workflows — they appear in the list too, but you'll rarely call them directly.
Pipelines: Scheduled Team Runs
Pipelines are predefined goal + preset combinations that can run on a schedule. The built-ins:
| Pipeline | Default Schedule | What it does |
|---|---|---|
memory-consolidation | Daily at 3am | Finds duplicate and conflicting memories |
daily-summary | Daily at 10pm | Summarizes the day's activity |
research-digest | Weekly, Monday 9am | Reviews all active projects |
Run one immediately, or put it on the calendar — both in plain language:
"Run the memory consolidation pipeline now"
→ crow_run_pipeline({ pipeline: "memory-consolidation" })
"Schedule the daily summary"
→ crow_schedule_pipeline({ pipeline: "daily-summary" })
"Run memory consolidation every Sunday at 2am"
→ crow_schedule_pipeline({ pipeline: "memory-consolidation", cron_expression: "0 2 * * 0" })Pipeline results are stored as Crow memories (tagged pipeline,automated), so the output lands in the same searchable knowledge base as everything else.
Watching It Work
Open Orchestrator in the Crow's Nest dashboard for a live timeline of every run: which agents dispatched, which model each used, token counts, durations, and any errors. It's the layer-down view — you don't need it for orchestration to work, but it's there when you're curious what actually happened.
Which Model Does It Use?
By default, the same provider configuration as Crow's AI chat (models.json). Two environment variables override the default for orchestration: CROW_ORCHESTRATOR_PROVIDER and CROW_ORCHESTRATOR_MODEL. Individual agents inside a preset can pin their own provider/model, so a team can mix local and cloud models in one run.
Next Steps
- Orchestrator architecture — presets, the MCP bridge, remote-instance tools, and internals
- Scheduling guide — how Crow's schedules work in general
- Context & Performance — how the orchestrator's tools fit your AI's context budget