Orchestration

A registry, a dispatch queue, a pipeline tracker, a session briefing, and operational runbooks. Wake up to a system that already knows the plan.

4 core files + 4 runbooks · System backbone

The problem

Without a system, you forget what you were working on. Tasks fall through cracks. You spend the first 10 minutes of every session figuring out where you left off. There's no queue, no priority, no staleness detection.

Most people track work in their heads, or in a notes app with no structure. That works for five items. It breaks at fifty. Orchestration is the structural fix: a set of files that hold state so your brain doesn't have to.

What gets encoded

Four core files wire the system together, plus operational runbooks that define pre-built agent sequences:

registry.json

Single source of truth for every work item. Projects, tasks, theses, backlog. Status tracking, goal mapping, timestamps. If it isn’t in the registry, it doesn’t exist.

dispatch.md

Active queue with decision types: auto, approve, review, input. Blocking relationships. Successor rules — when one task finishes, the next one queues automatically.

dashboard.md

Auto-generated pipeline tracker. Today’s priorities, staleness flags, what needs your input. Regenerated from registry every session so it’s always current.

briefing.md

Session-start briefing. What’s overdue, what’s ready, what changed. Generated by a Python script so you wake up to a system that already knows the plan.

Operational runbooks

Pre-defined agent sequences for common scenarios. Each runbook specifies trigger, agents, phases, handoffs, quality gates, time caps, and escalation paths.

Competitive Analysis

Triggered by a competitor move or scheduled refresh. Scout → Intel/Industry → Synthesizer → Writer. Duration: 1–2 sessions.

Morning Briefing

Daily at 7 AM (automated). Data Scout + Signal Monitor + Intel run in parallel, then Synthesizer assembles. Duration: ~5 minutes.

Artifact Production

On-demand PM artifact. Context Assembly → Producer + Skill → Quality Gates → Fix Loop. Duration: 1–2 sessions.

Incident: Metric Drop

Triggered by >10% WoW drop detected. Signal Monitor → Data Scout → Intel → Prime escalation. Duration: 1–3 hours.

Real example

Here's an actual work item from Agent Prime's registry — a task for this landing site:

shared/registry.json — Work Item

{
  "id": "TSK-014",
  "title": "Build landing site for open-source release",
  "type": "task",
  "status": "in_progress",
  "goal": "Matter",
  "parent": "PLN-009",
  "assigned_to": "builder",
  "created": "2026-03-08",
  "updated": "2026-03-10",
  "next_action": "Build layer pages"
}

Every item has a type, a status, a goal it serves, and a next action. The registry doesn't just track what exists — it tracks what happens next.

How to set it up

Start with registry.json — a JSON array of work items. Each item gets an ID, status, and goal. Add dispatch.md when you have enough items to need sequencing. The dashboard and briefing scripts are optional but powerful — they turn raw state into actionable summaries.

// registry.json — minimum viable item
{
  "id": "TSK-001",
  "title": "Describe the work in one line",
  "type": "task",        // task | project | thesis | backlog
  "status": "backlog",   // backlog | planning | in_progress | done
  "goal": "Matter",      // which life goal this serves
  "next_action": "What happens next"
}

The key insight: track work items in structured data, not in your head. JSON is machine-readable, so agents can query it, scripts can generate dashboards from it, and nothing falls through the cracks.

What you get

shared/registry.json

Single source of truth for all work items. Every agent reads this first.

~1,900 lines · 50+ items

prime/dispatch.md

Active queue with decision types and successor rules.

44 active items

prime/dashboard.md

Pipeline tracker. Auto-generated from registry each session.

Auto-generated

prime/briefing.md

Session-start briefing. Overdue, ready, and changed items.

Auto-generated

shared/runbooks/

4 operational runbooks: Competitive Analysis, Morning Briefing, Artifact Production, Incident Response.

Pre-defined agent sequences