How Camelia's agent loop works (model, tool calls, up to 6 turns)

When you send Camelia a message, she doesn't just reply once — she runs an agent loop that lets her take real actions and then keep reasoning until the work is done.

The loop

The cycle is: the model produces a response, that response may include tool calls, each tool call is executed (and logged), and the loop continues with the results fed back in — until she reaches a final answer. This repeat-and-refine pattern is what lets her chain steps: read a file, then act on it, then write a new document, then summarize.

Capped at six turns

The loop is bounded to a maximum of six model turns per request. This keeps responses from running away and ensures each request resolves in a predictable amount of work. If a task is genuinely large, she'll typically use a dedicated tool built for big jobs (for example, the review loop for substantial deliverables, or background CAD generation) rather than spinning the main loop.

Every action is real and logged

Each tool call acts on your real project data, and each is recorded as an Agent-to-Agent (A2A) log entry in your project's activity ledger. That means the loop is both functional and transparent: the actions she takes show up in the process trace and feed the conversation node graphs.

What you see

While the loop runs, Camelia streams progress. The final user-facing answer is her prose; the intermediate tool calls, approvals, and produced artifacts are surfaced separately in the process trace rather than buried in the reply. If a step needs your sign-off (like an external write), the loop pauses for your approval before continuing.