Agents
Four agents, each with one job. They pass JSON messages to each other. The model picks tools and writes plans, reviews, and explanations of failures. Code checks every tool call against an allow-list, runs only what a person approved, and sets each pass or fail.
Choose an agent
Choose an agent to see its job, what it reads and writes, its tools, and the checks on its output.
Who decides what
- The model decides which tool to call, what a plan contains, whether a plan is safe, and the most likely cause of a failure.
- Code decides whether a tool may run, whether a person approved it, and whether each check passed. A summary from the model never sets a pass or fail.