Multi-Task Coordination Protocol

Problem

Quan sends 3-5 requests rapid-fire, switches topics mid-thread. Serial processing creates 30-60+ second latency while he waits.

Solution

Parallel agent coordination:

1. Fast ack (<3 sec) with task list + ETAs
2. Spawn agents in parallel for each task
3. Results trickle back as they complete
4. Link related findings

---

Model Routing Rules

Haiku (2-10 sec)

Use for:

• Simple lookups (file exists? value in spreadsheet?)
• Quick confirmations
• Status checks
• Fast acks
• String searches

Examples:

• "Does file X exist?"
• "What's in cell A1?"
• "Check if service is running"

Sonnet (10-60 sec)

Use for:

• Standard searches (Drive, Gmail, files)
• Document summarization
• Code edits/writes
• Multi-step workflows
• API calls with processing

Examples:

• "Search Drive for X"
• "Summarize this meeting"
• "Create script for Y"
• "Fetch emails and triage"

Opus (60-300+ sec)

Use for:

• Deep strategic analysis
• Complex architecture decisions
• Multi-document synthesis
• Long-term planning
• Business strategy

Examples:

• "Analyze cash flow impact across 6 scenarios"
• "Design RAG architecture for Phase 5"
• "Synthesize 10 meetings into quarterly review"

---

Coordination Flow

1. Request Arrives

User: "Search Drive for X, check if Y is ready, analyze Z impact"

2. Fast Ack (Haiku, 2 sec)

✅ Got 3 tasks:
1. Drive search for X (~20s)
2. Check Y status (~5s)
3. Z impact analysis (~2min)

Working...

3. Spawn Parallel

• Task 1 → Sonnet spawn
• Task 2 → Haiku spawn
• Task 3 → Opus spawn

4. Results Stream In

✅ Y status: Ready (completed 2s ago)

✅ Drive search: Found 3 matches for X:
[results]

✅ Z impact analysis: [detailed findings]

---

Task Classification Logic

Triggers:

• Multiple questions in one message

• 1 tool call needed

• 10 sec expected completion time

Classifier checks:

1. How many distinct requests?
2. Can they run in parallel?
3. What model for each?
4. Any dependencies?

Decision tree:

Single simple task → Just do it (current behavior)
Single complex task → Spawn once, report when done
Multiple tasks → Fast ack + spawn per task

---

Implementation Files

Task Queue:

• tools/task-coordinator.py - Main coordination logic
• tools/task-classifier.py - Request → model mapping

State:

• data/task-queue.json - Active task tracking
• Per-task status in spawn session metadata

---

Quan's Interaction Pattern

Observed:

• Sends 3-5 requests rapid-fire
• Switches topics (Fathom link → rebuild → multi-model → Sheets)
• Doesn't wait for answers before next question
• Wants continuous feedback, not batch reports

Optimized for:

• Low initial latency (fast ack)
• Parallel execution (no blocking)
• Streaming results (report as ready)
• Context linking (connect related findings)

---

Testing Protocol

1. Wait for next multi-request batch from Quan
2. Fast ack with task breakdown
3. Spawn parallel agents
4. Measure:
   • Time to first ack
   • Time to first result
   • Time to all results complete
   • Accuracy of ETAs

---

Created: Feb 13, 2026
Status: Building implementation