Function execute

Execute an agent task in a loop until exit is called. Provides planning, progress tracking, and controlled exit functionality.

The agent runs until the model calls the built-in exit tool, the maxIterations limit is reached, or an abortSignal is triggered.

Message injection

The messages array is used directly (not copied), so you can push new messages onto it at any point while the agent is running. They will be included in the context at the start of the next iteration:

const messages = [{ type: 'user', text: 'Perform the task.' }]

const stream = execute({ client, messages, tools })

// inject a user message or system notification mid-run:
messages.push({ type: 'user', text: 'Also handle edge case Y.' })
messages.push({ type: 'context', text: 'System: disk usage at 90%.' })
Copy

The agent also appends its own bot responses to the same array as each iteration completes, so messages reflects the full conversation history.