Skip to main content

Core concepts

Agent​

An agent is a graph of nodes plus its configuration and rules. You build it on the canvas, and it is saved as a versioned snapshot — editing an agent creates a new version rather than overwriting the last one.

An agent is either a draft or published. Drafts can only be tested from the dashboard; publishing unlocks API access.

Run​

A single execution of an agent. Every run:

  1. Is queued as a message rather than executed in the web request.
  2. Is picked up by a worker process, which loads the agent version, sorts the nodes topologically, and executes them in dependency order.
  3. Emits an event per node (node.start, node.complete) as it goes.
  4. Ends as COMPLETED or FAILED, with credits charged either way.

Because runs are queued, an API call returns immediately with a runId instead of blocking until the agent finishes.

Trace​

The per-node record of a run — input, output, duration, token counts, and any error. Traces stream live over SSE while a run is in progress, so you watch execution rather than waiting for a final result.

Credits​

The billing unit. Credits are charged for tool calls only — LLM tokens are billed by your own provider under BYOK. See Billing.

Rules​

Typed, versioned instructions attached to an agent (identity, mission, prohibitions, tone, edge cases) that compile deterministically into the system prompt every LLM node inherits. See Agent rules.

Tools​

Reusable capabilities an agent node can call — nine built-ins plus custom Python that runs in a WebAssembly sandbox. Managed under Dashboard → Tools. See Tools.

Pipelines​

Scheduled data syncs, modelled as reusable Source → Destination → Connection objects with incremental cursor-based loads and SQL transforms. They run on the same worker and event infrastructure as agent runs. Managed under Dashboard → Data. See Data pipelines.

Security scans​

Static analysis of an agent's graph plus dynamic probes that attack the agent through its normal execution path, scored and mapped to the OWASP LLM Top 10 and NIST AI RMF. Findings explain what the weakness is, how it manifests in your specific agent, and what to change. Run from Dashboard → Security. See Security scanning.