Agent Coordination

DOSI defines how human and AI collaborate — and how the AI dispatches parallel work.

The Dispatch Model

Supervisor (human)
    │
    ├── sets objectives ──────→ confirmed before work starts
    ├── injects ideas  ───────→ captured immediately, triaged later
    ├── approves/steers ──────→ decisions that need human judgment
    └── reads status   ←──────  OCD file is the live report

Orchestrator (AI)
    │
    ├── dispatches parallel agents for independent tasks
    ├── sequences dependent tasks
    ├── captures and logs all decisions
    ├── queries knowledge network for context
    └── updates OCD + STATE continuously

The human doesn't micromanage. The AI doesn't go rogue. The OCD file is the shared contract.

Parallel Work Rules

1. Independent tasks → parallel. Two tasks that don't share files or state run concurrently.
2. Dependent tasks → sequential. If task B needs task A's output, wait.
3. Review after major steps. Pause and check before continuing past a logical boundary.
4. Never duplicate work. If an agent is researching X, don't also research X.

Status Reporting

The OCD file IS the live status report. At any point, the supervisor reads it to see:

• Actions Taken — what's done (checked items)
• In Progress — what's happening (unchecked items)
• Issues & Blockers — what's stuck
• Idea Capture — raw ideas logged during the session

No separate status meetings. No "where are we?" questions. The OCD file answers everything.

Multi-Agent Sessions

For complex sessions, the orchestrator dispatches specialized agents:

Orchestrator
    ├── Agent 1 → scaffold product-engine (Go project)
    ├── Agent 2 → repair admin-portal (fix dosi/)
    └── Agent 3 → update HQ state

Each agent works in its own directory. No shared state. Results merge at the end.

Agent Reviews

At session close, different perspectives review the work:

Perspective	What It Reviews
Orchestrator	Session management, git, architecture
Pipeline	Processing health, model performance
Business	Strategic alignment, market readiness
Review	Code quality, conventions, tech debt

Not every session needs every perspective. Use what's relevant.