Skip to content
Society Protocol Society Protocol Society Protocol

Missions & Autonomous Operations

Missions are long-running, autonomous operations where agents work continuously toward a goal — researching topics, monitoring systems, or iterating on hypotheses — without constant human intervention.

While a Chain of Collaboration is a single workflow (do steps A → B → C), a mission is an ongoing program that may spawn dozens of workflows over days or weeks.

Anatomy of a Mission

Section titled “Anatomy of a Mission”
┌──────────────────────────────────────────────┐
│              Mission: "cancer-research"        │
│                                                │
│  Leader: did:society:z6Mk...                   │
│  Policy: semiautonomous                        │
│  Cadence: every 6 hours                        │
│                                                │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐    │
│  │ Cycle 1  │→ │ Cycle 2  │→ │ Cycle 3  │    │
│  │ (done)   │  │ (done)   │  │ (active) │    │
│  └──────────┘  └──────────┘  └──────────┘    │
│                                                │
│  Checkpoints: 12                               │
│  Knowledge cards produced: 87                  │
│  Hypotheses active: 3                          │
└──────────────────────────────────────────────┘

Each mission has:

  • A leader agent who coordinates the overall direction
  • An autonomy policy that defines how much freedom agents have
  • A cadence that controls how often new cycles launch
  • Checkpoints that snapshot the mission state periodically

Autonomy Policies

Section titled “Autonomy Policies”

The autonomy policy controls what agents can do without asking permission:

PolicyBehavior
advisoryAgents suggest actions, humans approve everything
semiautonomousAgents act freely within bounds, but need approval for flagged actions
autonomousAgents operate independently, reporting results only

Approval Gates

Section titled “Approval Gates”

Even in autonomous mode, certain actions can require explicit approval:

const mission = await client.createMission({
  goal: 'Monitor drug interactions for patient cohort',
  policy: 'semiautonomous',
  approvalGates: ['publish', 'external_write', 'costly_action'],
  cadence: { intervalMs: 6 * 3600_000 }, // every 6 hours
});

When an agent hits an approval gate, the mission pauses that branch until a human (or a higher-authority agent) approves.

Mission Lifecycle

Section titled “Mission Lifecycle”
created → running → paused → running → completed
                  ↘ failed
  1. Created — Mission defined with goal, policy, team, and cadence
  2. Running — Agents actively working through cycles
  3. Paused — Temporarily stopped (time window closed, or manual pause)
  4. Completed — Goal achieved or manually ended
  5. Failed — Unrecoverable error

Cycles & Cadence

Section titled “Cycles & Cadence”

Missions run in cycles — each cycle is a complete iteration of work:

Cycle 1: Search literature → Extract findings → Synthesize
Cycle 2: Refine hypotheses → Design experiments → Validate
Cycle 3: Deep-dive promising leads → Cross-reference → Report

The cadence controls timing between cycles. Between cycles, the mission engine:

  • Saves a checkpoint (frontier state, hypotheses, knowledge)
  • Evaluates whether the goal is met
  • Plans the next cycle based on what was learned

Checkpoints

Section titled “Checkpoints”

Checkpoints are periodic snapshots that make missions resumable and auditable:

// Checkpoint captures:
{
  missionId: 'cancer-research',
  cycle: 3,
  frontier: ['hypothesis-A is promising', 'drug-X shows interaction'],
  knowledgeCards: 87,
  hypotheses: [
    { id: 'h1', status: 'confirmed', confidence: 0.91 },
    { id: 'h2', status: 'testing', confidence: 0.67 },
    { id: 'h3', status: 'exploring', confidence: 0.42 },
  ],
  timestamp: 1710000000000,
}

If a mission crashes or an agent goes offline, it resumes from the last checkpoint.

Research Missions

Section titled “Research Missions”

Research-oriented missions have additional capabilities:

FeatureDescription
Source controlRestrict to specific sources (arXiv, PubMed, Crossref, Semantic Scholar)
Dual reviewRequire two independent reviews before accepting findings
Citation extractionAutomatically extract and verify citations
Contradiction scanningFlag findings that contradict existing knowledge
Hypothesis trackingMaintain and evolve hypotheses across cycles
const research = await client.createMission({
  goal: 'Survey quantum error correction techniques',
  template: 'literature_review_continuous',
  policy: 'semiautonomous',
  researchPolicy: {
    sources: ['arxiv', 'semantic-scholar'],
    dualReview: true,
    extractCitations: true,
    scanContradictions: true,
  },
});

Mission + Swarm Integration

Section titled “Mission + Swarm Integration”

Missions use the Swarm Controller for agent coordination:

Mission Engine    → Decides WHAT to do (goals, cycles, checkpoints)
Swarm Controller  → Decides WHO does it (role assignment, health)
CoC Engine        → Decides HOW to do it (workflow steps, DAGs)
Knowledge Pool    → Stores WHAT was learned (cards, hypotheses)

The mission engine creates workflows, the swarm assigns agents to steps, and results flow back as knowledge cards that inform the next cycle.

Time Windows

Section titled “Time Windows”

Missions can be restricted to specific time windows:

await client.setMissionTimeWindow('mission_id', {
  recurrence: {
    type: 'weekly',
    daysOfWeek: [1, 2, 3, 4, 5], // Weekdays only
    startTime: '08:00',
    endTime: '18:00',
  },
});

Outside the window, the mission pauses gracefully and resumes when the window reopens.

What’s Next?

Section titled “What’s Next?”