Agent Loop

The agent loop is the core execution cycle that drives all agent behavior.

Loop diagram

Loop stages

Stage	Description	Hook points
1. Input	Receive user instruction	—
2. Context	Load skills, tools, context files	beforeAgentStart, transformContext
3. Prompt	Build system prompt with tools	beforeProviderRequest, beforeProviderPayload
4. Request	Call LLM with streaming	—
5. Parse	Extract tool calls from response	afterProviderResponse
6. Execute	Run tool calls	beforeToolCall, afterToolCall
7. Repeat	Feed results back to LLM	prepareNextTurn, shouldStopAfterTurn
8. Output	Render final response	—

Error handling

The backend classifies errors into categories:

Category	Retryable	Action
context_full	No	Compact session
rate_limit	Yes	Backoff retry
transient	Yes	Backoff retry
client	No	Report to user
poisoned_history	No	Compaction
unknown	No	Report to user

Iteration limits

{
  "maxIterations": 10,
  "maxTokens": 8192
}

The loop terminates when:

• The agent produces a final response (no tool calls)
• maxIterations is reached
• maxTokens is exceeded
• The user cancels