- A unique ID
- An API key
- Its own credit balance
- Its own policy (caps, allowed capabilities)
- Its own audit trail
Why Agents Exist
Agents provide isolation boundaries. When Agent A exhausts its budget, Agent B continues operating. When Agent A’s key is compromised, you revoke one key—not your entire system. Without per-agent isolation:- One runaway loop drains your entire balance
- One compromised key exposes all capabilities
- Cost attribution is impossible
Enforcement Behavior
- Every API call requires a valid agent key
- Invalid or revoked keys return
401 Unauthorized - Agent keys are scoped to their policy—they cannot exceed their caps regardless of account balance
- Kill switch on an agent immediately blocks all calls for that agent
Common Patterns
One Agent Per Workload
One Agent Per User (Multi-tenant)
Separate Environments
Common Mistakes
| Mistake | Consequence |
|---|---|
| Using one agent for all workloads | No isolation, no attribution |
| Sharing keys across environments | Production incidents from dev mistakes |
| Not rotating keys after team changes | Security exposure |
| Creating agents without caps | Unbounded spend risk |