Core concepts

This page defines the nouns and verbs used throughout the rest of the docs. Refer back to it whenever a term is unclear.

Agent

The thing you build: a configuration of nodes, tools, and a system prompt that determines what it's allowed to do and how it behaves. You build an agent with the conversational builder or the canvas builder.

Agent version (release)

Agents are versioned. As you edit an agent, changes accumulate in a draft; when you're ready, you release a version. A released version is immutable and is what gets deployed, scheduled, or rolled back to.

Run

One execution of an agent version, from trigger to completion (or failure, or a pending approval). A run is not a single model call — it's a full pass through the agentic loop, bounded by a step limit.

Step

One iteration of the agentic loop within a run: the model chooses a tool, the tool's arguments are validated, the tool executes, and the result is fed back to the model. Each step is recorded in the run's execution trace.

Tool

A capability the model can invoke during a run. Available tool types include:

• code_execution — sandboxed JS (QuickJS WASM) or Python (Pyodide) execution
• api_call — outbound HTTP with SSRF protection
• knowledge_search — RAG search over your knowledge base with citations
• connector_action — calling a connected third-party service
• MCP tools — tools exposed by a connected Model Context Protocol server
• subagent delegation — handing a sub-task to a specialized subagent

Connector

A configured connection to a third-party service (Google Workspace, Slack, Salesforce, and others) that an agent can call via connector_action. Connectors authenticate via one-click OAuth or a pasted API key/token; credentials are encrypted and org-owned.

Trigger

What starts a run automatically: an inbound webhook (HMAC-verified), authenticated polling, a cron schedule, or an inbound channel such as email, Telegram, WhatsApp, a web form, or an embeddable widget.

The agentic loop

The multi-turn tool-calling loop every run executes:

• Bounded steps — a run stops after a fixed maximum number of steps (default 8), win or lose.
• Arg validation — each tool call's arguments are checked against a JSON Schema before execution; validation errors are fed back to the model instead of crashing the run.
• Approval gate — destructive tool calls pause the run and wait for a human to approve before executing.
• Refuse rather than guess — the system prompt instructs the model to decline rather than fabricate an answer when it doesn't have enough information.

The loop works the same way whether the underlying model is OpenAI (function-calling) or Anthropic (tool-use).

Organization / roles

Every agent, run, connector, and credential belongs to an organization. Access within an organization is controlled by role — every query and mutation is scoped to your organization and checked against your role before it runs.