Agents

Hosted agent harnesses that run inside your boundary.

An Allocate agent is composition as data: a model, a harness, a sandbox, a tool list, and a system prompt, declared as configuration rather than code. The platform runs the loop; you decide what the agent is.

The pieces

• Model: any catalog model or Route, with per-agent thinking level.
• Harness: the reasoning loop behind a driver contract; harnesses are selectable per agent and versioned, so you can compare them on the same traffic.
• Sandbox: every agent runs in its own isolated sandbox with a warm window. Nothing an agent does touches your systems without a grant.
• Tools: each tool declares where it executes. Tools that need credentials run at the platform edge, so provider, database, and API keys never enter the sandbox.
• Memory: durable per-agent memory that survives sandbox sleep, curated by the agent itself, plus per-agent skills: procedural playbooks the agent saves and reuses.

Write paths are human-gated

An agent can propose a write: a pull request, a platform action, an outbound message. Only a verified human click performs it. Proposals are dry-run first, show their real effects, and apply exactly once. There is no configuration in which a model turn executes a write directly.

The fleet

Every agent turn is recorded as a run, failures included, and every run is visible in your dashboard alongside its Meter data. External harnesses that call the Gateway appear in the same fleet view as platform-managed agents, because the capture happens at the wire.