The agent builder tabs

Basics. The agent's name, its default app, description, avatar and tags. An app's Router reads the description when it chooses between agents.

Instructions. The system prompt. It is sent with every model call unless a workflow node overrides it. Keep it specific: audience, tone, what to do when information is missing, when to use tools.

Model. Provider, model, temperature, max output tokens, the context limits (model context limit, max input tokens and a safety margin), retry count and a fallback model. An LLM node in the workflow uses these unless a model is set on the node itself. See Models and providers.

Tools. Attach tools from the organisation library: Add tools lists them by toolbox, so you can pick a whole toolbox, tools inside it, or tools in no toolbox. Each picked tool is attached on its own; with a single tool picked you can also customise it for this agent (alias, description, config, approval). Attached tools are inherited by the agent's toolbox node. Tool settings such as authentication, safety level and limits live on the tool itself (Tools).

Knowledge. Attach knowledge bases. The knowledge retrieval node uses them by default.

Memory. Memory mode and defaults: off, remember in this chat, remember for this agent, remember this user; plus the namespace and read/write defaults (Memory).

Workflow. The graph editor: nodes, edges, conditional edges, per-node settings, validation, and versions (Workflows: nodes, edges and the default graph).

Advanced. Structured output is an optional output schema: when it is set, the workflow validates the answer as JSON against the schema and can retry or repair it (validation node, Workflows: nodes, edges and the default graph). Runtime settings hold the timeout, the maximum cost per run, max iterations and whether answers stream.

Max iterations is the agent's tool budget per turn: how many rounds of tool calls it may make before it must answer. The model is told the budget before every call and warned on its last round; when the budget is spent it gets one more turn without tools to answer with what it has, saying what it could not check. Size it by the work: a few rounds for lookups, more for agents that page through large results.

Logs. Runs of this agent with every step, prompt, tool call and retrieval (Testing, logs and observability).

Publish. Publish the draft, see published versions, restore an earlier version.

Approval is set on each tool, or for one agent when you attach a single tool. Input and output safety rules are set on the app's Chat (in) and Chat (out) nodes. Both are described in Guardrails, approvals and safety.