Agents

An agent is an autonomous LLM-powered process that receives a task, reasons about how to accomplish it, uses tools, and delivers a result.

The ReAct Loop

Every agent runs a ReAct (Reason + Act) loop. On each iteration:

• Context is assembled: system prompt, skills, memory, tool definitions, conversation history
• The LLM generates a response — either text or tool calls
• If tool calls: execute them, append results, loop again
• If text: stream to client, conversation is complete

The loop runs up to maxIterations (default 25) before stopping. Each iteration is metered for billing.

Configuration

• systemPrompt — the core instructions. Injected at the start of every context window
• model — which LLM to use (routed through AI Gateway)
• temperature — controls randomness (0.0–1.0)
• maxIterations — loop limit per conversation turn
• enabledTools — which tools the agent can access (filtered by plan tier)

Subagents

Agency-tier agents can spawn subagents — background agent instances with their own context windows but access to the same workspace. Subagents are useful for parallel research, independent sub-tasks, or divide-and-conquer strategies.

Depth is limited to 3 levels to prevent infinite recursion. Subagent iterations are capped at 15.

Streaming

Responses stream via Server-Sent Events (SSE):

• event: text — streamed text chunk
• event: tool_start — tool call initiated
• event: tool_result — tool call completed
• event: done — conversation turn complete with stats
• event: error — error occurred