Skip to main content

Worker management

Define, run, and govern AI and human workers with full cost and quality visibility.

Raytio Worker management describes every worker - human or AI - and everything needed to spin one up: role, template, system prompt, model stack, and toolset. It records each dispatch as a structured run, evaluates the quality of the output, checks it against guardrails, and tracks its cost against budgets - governed work with the same access control and multi-tenancy guarantees as the rest of the platform, and a human kept in the loop wherever the work is high-risk.

Capabilities

Definitions as templates, workers as instances

A definition captures role, system prompt, model preferences and tool access once. Workers are instances of that definition, and can override settings without touching the template.

Three-tier model stack with failover

Providers, models and deployment availabilities are configured separately. Definitions hold an ordered preference list, and dispatch falls back to the next choice when a model is unavailable or rate-limited.

Tool access by allowlist

Workers connect only to explicitly granted tool servers, narrowable to specific tools within a server.

Agent runs as an observability record

Every dispatch is recorded as a structured run.

  • The worker, and the model and provider actually used.
  • Tokens, cost and duration.
  • Branch and merge-request links.
  • Final status and a step-by-step trace.
Quality evaluation after each run

Rule-based, model-graded and statistical scorers assess completed work on a 0-1 scale. Scoring runs asynchronously, so it never blocks delivery. A different model can be used for cross-model validation.

Guardrails before and after dispatch

Pre-dispatch checks scan the prompt and budget before any cost is incurred. Post-dispatch checks inspect the output. Each check is rule-based or model-graded, at warn or block severity.

Budgets and circuit breakers

Budgets cascade from role to definition to instance, and are checked before dispatch. Budget periods can be daily, weekly, monthly or per-run. A circuit breaker skips providers with too many recent failures.

Human-in-the-loop approval gates

High-risk work pauses and waits for a human decision, with configurable reminder and escalation timeouts. The system never auto-approves.

Role memories

Persistent, typed notes attach to a role, so a worker carries forward past decisions rather than starting cold. The note types are feedback, lesson, reference and project.

Orchestration

A supervisor decomposes an epic into dependent child work items. The items are dispatched in parallel where dependencies allow, then rolled up into a summary on completion.

Notable details

Run states

The run states are RUNNING, SUCCESS, FAILED, TIMEOUT, BUDGET_EXCEEDED and INVALID_OUTPUT. Every run ends in an explicit state.

Credentials stay separate

Provider API keys are held encrypted in the platform's merchant authentication system, never in worker configuration.

Dependency-aware scheduling

Three factors determine what dispatches next: standard dependency types, concurrency limits, and same-repository serialisation.

Part of Agents · Read the documentation →

Verified customers, protected data, and agents you can hold to account.