Orchestrator

The orchestrator is the top-level AI agent in your lab. It is your primary interface for directing research. You talk to the orchestrator in natural language, and it translates your intent into structured tasks distributed across the agent team.

How It Works

You: "Run an MCMC chain with the Planck 2018 likelihood, 100K samples"
                    │
          Orchestrator (Opus)
           ┌────────┼────────┐
           ▼        ▼        ▼
     Parse intent  Check    Select
     + parameters  compute  agent
           │        │        │
           ▼        ▼        ▼
     Create EXP   Allocate  Assign to
     with config  H100 pod  Research Lead

The orchestrator:

1. Parses your request into a structured task
2. Determines the reasoning level required
3. Routes to the appropriate agent
4. Allocates compute resources
5. Monitors execution and reports back

Chat Interface

The orchestrator chat works like any messaging interface, but with superpowers:

• Natural language — No special syntax required. Just describe what you want.
• Context-aware — The orchestrator knows your lab's history, experiments, and current state.
• Multi-turn — Follow-up questions refine the task. "Actually, use 200K samples" works naturally.
• Proactive — The orchestrator surfaces blockers, suggests next steps, and flags idle resources.

Example Conversations

Running an experiment:

You: Run the birefringence analysis on the Planck EB cross-spectrum
Orch: I'll set up the NaMaster birefringence experiment.
      - Pod: H100 (estimated 2 hours)
      - Config: planck_eb_cross.yaml
      - Agent: Research Lead
      Starting now. I'll report back when QC completes.

Checking status:

You: What's running right now?
Orch: 3 experiments active:
      - EXP-054: MCMC base chain (78% complete, ~40 min remaining)
      - EXP-055: f_NL Fisher forecast (running, 2 hours elapsed)
      - EXP-056: Figure generation (queued, waiting on EXP-054)
      2 pods active. GPU utilization: 94%. Cost today: $12.40.

Strategic direction:

You: The ECH routes are all closed. What bounce models can still work?
Orch: Based on the 14 structural barriers we've documented, three
      bounce models bypass ECH-specific closures:
      1. Quintom-B (w-crossing) — favored at 2.3σ in our MCMC
      2. Matter bounce — parameter-free f_NL prediction testable by SPHEREx
      3. Cuscuton — avoids ghost instability via spatial gradient terms
      Shall I prioritize experiments for any of these?

Standups

The orchestrator runs 3x daily standups (morning, midday, evening) that summarize:

• What was accomplished since the last standup
• What is currently running
• What is blocked and needs your input
• Recommended next actions

# Trigger a standup manually
hubify standup

# View standup history
hubify standup list

Escalation

The orchestrator escalates to you when:

• An experiment fails and the cause is ambiguous
• A QC gate fails and the agent cannot determine why
• Two agents disagree on a result (conflicting reviews)
• Budget limits are approaching
• A task requires a decision outside the agent's authority

Escalations appear as highlighted messages in the chat and as notifications in Captain View.

Configuration

# Set the orchestrator model
hubify agent update orchestrator --model claude-opus

# Set standup schedule
hubify agent update orchestrator --standup-schedule "8am,12pm,6pm"

# Enable auto-scheduling (orchestrator picks next experiment when pods are idle)
hubify agent update orchestrator --auto-schedule true