Multi-agent teams

By the end, each agent in your team has its own key and budget, your policy treats them differently, and you can stop one without stopping the rest.

Before you start

An org admin key.

1. One agent per role

deedglass agent create researcher --label team=support --calls-per-minute 60
deedglass agent create refunder --label team=support --label money=yes --spend-per-day 50000 --currency usd
deedglass agent keys create refunder --mode test

Each agent gets its own key. Never share one key across agents: the Ledger, budgets and Stop work per agent.

2. Write rules per label

  - id: refunds-for-money-agents
    match: { tool: stripe.create_refund }
    when: agent.labels.money == "yes" && args.amount <= 5000
    decision: allow
    verify: L1

3. Pass the person along

When an agent acts for a user, send Deedglass-On-Behalf-Of (on_behalf_of= in Python, onBehalfOf in TypeScript). The deed records it, and Insights watches for bursts of new on-behalf-of users.

4. Stop one

deedglass agent stop refunder --reason "…" stops only the refunder. The researcher keeps working.

What it doesn’t do

Deedglass doesn’t orchestrate agents or pass messages between them. It governs each one’s tool calls.
An agent that calls another agent’s API with that agent’s key is recorded as that agent. Keep keys apart.

Next