Built for How Agents Work

// Write state
await agent.state.write("conversation", {
  messages: [...],
  context: {...},
  lastAction: "completed_task_123"
});

// Read state (next session, different instance)
const state = await agent.state.read("conversation");
// Picks up exactly where you left off

// Opus documents the workflow
await agent.workflows.create({
  name: "customer-onboarding",
  model: "opus",
  definition: `
    # Customer Onboarding
    1. Check CRM for existing record
    2. If new: Create contact with...
    3. Send welcome email using...
  `
});

// Haiku executes it perfectly
const workflow = await agent.workflows.get("customer-onboarding");
// Context is perfect, no intelligence needed

// Opus hands off to Haiku
await agent.handoffs.create({
  to: "haiku-executor",
  workflow: "customer-onboarding",
  context: {
    customer: {...},
    state: "ready_for_email"
  }
});

// Haiku picks it up
const handoff = await haikuAgent.handoffs.pending();
await haikuAgent.handoffs.accept(handoff.id);
// Continues work with full context

// Low confidence? Escalate.
if (confidence < 0.7) {
  await agent.escalate({
    reason: "Unusual refund amount",
    context: { customer, amount },
    suggestedAction: "Approve manually"
  });
}

// Human reviews in dashboard
// Agent continues when resolved

// All actions automatically logged
agent.state.write(...)   // → audit log
agent.handoffs.create(...) // → audit log
agent.escalate(...)      // → audit log

// Query the audit trail
const actions = await agent.audit.query({
  agent: "my-agent",
  since: "2024-01-01",
  type: "handoff"
});