Agent fleet on Relixr

A reference layout for running many autonomous agents behind one Relixr gateway, each with its own key, budget, and usage trail.

If you're running more than a handful of autonomous agents — crawlers, coding agents, support bots, batch jobs — you need answers to three questions for each one: which model is it allowed to call, how much can it spend, and what did it actually do. This architecture gives you all three without building your own control plane.

The building blocks

• One key per agent. Each agent gets its own key with scoped permissions — model allowlist, route allowlist, and a spend ceiling. If an agent misbehaves, you revoke one key, not the whole fleet.
• Budgets as a backstop, not a dashboard afterthought. Set the ceiling when you provision the key, before the agent runs, so a runaway loop fails a request instead of draining the account.
• Attribution by key, not by inference. Every response and every usage row carries the key id, so you can answer "what did Agent C cost this week" with a filter, not a spreadsheet.
• Shared routing and guardrails. Because every agent goes through the same gateway, a routing rule change or a new guardrail applies to the whole fleet the moment you save it — you don't redeploy each agent.

Applying this to your fleet

1. Create one key per agent (not one shared key for all agents) — see Create your first API key.
2. Scope each key to only the models and routes that agent needs. See API key scopes.
3. Set a per-key spend ceiling sized to that agent's expected workload.
4. Use the usage analytics view, filtered by key, as your per-agent cost and activity report.