Agents & Ara

import { ArqeraClient } from '@arqera/sdk';

const client = new ArqeraClient({ apiKey: 'ak_...' });

// AUTO action — executes immediately
const report = await client.ara.execute({
  action: 'report.generate',
  params: { type: 'monthly_summary', period: '2026-01' },
});

// SOFT action — 30-second undo window
const ticket = await client.ara.execute({
  action: 'ticket.create',
  params: { title: 'Review Q1 metrics', priority: 'medium' },
});

// HARD action — requires approval
const pending = await client.ara.execute({
  action: 'email.send',
  params: {
    to: '[email protected]',
    subject: 'Partnership Proposal',
    body: 'We would like to discuss...',
  },
});
// Returns { status: 'pending_approval', approval_id: 'apr_...' }

// Approve the pending action
await client.ara.approve(pending.approval_id);

// Physarum routing: reinforcement learning for model selection
//
// For each request:
//   1. Score all available providers by conductance
//   2. Apply Bellman cost: C = risk * 0.4 + latency * 0.3 + cost * 0.3
//   3. Select lowest-cost provider that meets quality threshold
//   4. Execute request
//   5. Update conductance: success → reinforce, failure → decay
//
// Conductance dynamics:
//   dP/dt = 0.3 * |flow| - 0.05 * P
//   Success → conductance += 0.3 * |1/latency|
//   Failure → conductance *= 0.95
//
// Fallback chain:
//   Primary → Secondary → Tertiary → OpenRouter (last resort)
//
// Example routing decision:
//   Task: "Analyze quarterly revenue trends"
//   → Category: reasoning + analysis
//   → Top candidates: DeepSeek (conductance: 8.2), Claude (7.9), Groq (6.1)
//   → Selected: DeepSeek (lowest Bellman cost for this task type)
//   → Fallback: Claude → Groq → OpenRouter