What to set first
Instructions
The system prompt. Up to 10,000 characters at creation, and up to 50,000 once you edit it afterwards. This is the single biggest lever on behaviour — see Instructions. The same view carries an optional starter message, up to 2,000 characters: the agent’s opening line. Leave it blank and the agent waits for the user.Model
Smart routing by default: the platform scores each request locally and picks a model, and the agent picks up new models as they ship. Pin a specific model when you need reproducibility — names are in the model catalog. See Model routing.Capabilities
Bind built-ins, plugins, and connectors from the Capabilities view. Start narrow. See Built-in capabilities.Autonomy
New agents default to Supervised: Read-tier tools run freely, Act and Privileged tools pause for approval unless allowlisted. Decide deliberately whether that is right before the agent goes anywhere a human is not watching — on unattended surfaces a would-be pause becomes a denial, not a prompt. See Autonomy and approvals.What to leave alone at first
Under Advanced:- Tool presentation — Adaptive resolves per session from how many tools are bound. Pin a mode only when you have a reason.
- Tool choice and reset after first call — the defaults prevent tool loops.
- Structured output — only when a machine, not a person, reads the answer.
- Generative UI — off unless the client renders A2UI. See Generative UI.
Attaching the rest
Testing
Open the agent in the Playground. The agent header has an “Open in Playground” button when you have execute authority on it. Add?debug=1 to the Playground URL to see the assembled system prompt and the
tool and skill snapshots frozen for that run. That is the fastest way to find out
why an agent did not use a tool you thought it had.
Approvals render inline in the Playground with three answers: allow once, allow
for this thread, deny.