Agents

Agent Settings

Fine-tune how your agent executes — planning behavior, timeouts, streaming, and tool execution backends.

Default Configuration

json

{
  "settings": {
    "planningMode": "initial_plan",  // off | initial_plan | per_step_plan
    "emitPlanSummary": false,        // Include plan summary in output
    "timeout": 120,                  // Max execution time (seconds)
    "streamFinalOnly": true,         // Only stream the last step
    "defaultToolBackend": "client"   // client | webhook
  }
}

Settings Reference

planningModestring

Controls whether the agent creates a plan before executing steps.

"off"

No planning. Steps execute immediately.

"initial_plan"

Create a plan before the first step, then execute. Default.

"per_step_plan"

Re-plan before each step. More thorough but slower and more expensive.

emitPlanSummarybooleandefault: false

When enabled, the agent emits a short plan summary as part of its output before executing steps. Useful for transparency and debugging.

timeoutnumberdefault: 120

Maximum execution time in seconds (1–600). If the agent exceeds this, execution is aborted and a partial result is returned.

streamFinalOnlybooleandefault: true

When true, only the final step's output is streamed to the client. Intermediate steps run silently. Set to false to stream all steps (useful for debugging or showing the agent's thought process).

defaultToolBackendstringdefault: "client"

Default execution backend for custom tools that don't have an explicit webhookUrl.

"client"

Return tool_calls to the client for execution.

"webhook"

Attempt to execute server-side via webhook.

Router Assignment

Each agent is linked to a router that handles LLM model selection for all prompt steps. The router automatically picks the best model based on your optimization preferences (cost, latency, quality).

Assign a router in Agent Editor → Settings → Router. If no router is assigned, the agent will use your account's default router.

Active / Inactive Status

Agents have an isActive flag. Inactive agents cannot be called via API — requests return a 403 error. Toggle the status from the agent editor header.

When to use each planning mode

off — Simple single-task agents (chatbots, translators).
initial_plan — Multi-step agents that benefit from upfront reasoning. Best default.
per_step_plan — Complex agents where each step depends heavily on the previous result. Doubles LLM calls.

Next: Agent API Reference

Full API docs with code examples and error codes.