Run an agent safely
Ce contenu n’est pas encore disponible dans votre langue.
Run an agent safely
Section titled “Run an agent safely”In Agent mode, the backend Agent Runtime talks to the model and calls tools. Chat shows the run as it happens, including approvals, questions, tool activity, and the final answer.
Before you start
Section titled “Before you start”You need:
- an Agent profile;
- a configured provider/model;
- the tools that agent is allowed to use.
If needed, first follow Configure an agent.
1. Select the Agent profile
Section titled “1. Select the Agent profile”Open Chat and select the Agent profile you want to use.
The selected Agent profile carries its provider/model defaults, system prompt, tool policy, skills, MCP assignments, and optional default context.
2. Send the task
Section titled “2. Send the task”Write the task normally and send it.
As the run progresses, Chat can show tool calls, skills, delegation, and other runtime events. Tool details stay in compact disclosures instead of being mixed into the normal assistant message.
3. Handle an approval request
Section titled “3. Handle an approval request”When Runtime policy requires confirmation, Chat shows an Approval card.
Review:
- the action title and summary;
- the risk label;
- Arguments;
- any file/edit preview shown for the action.
Then choose:
- Approve — run this exact pending request;
- Reject — decline this request;
- Cancel — send the Runtime cancel decision for the pending request.
Approve, reject, and cancel apply to the exact pending tool call. Chat does not execute a separate copy of the action.
4. Answer a Runtime question
Section titled “4. Answer a Runtime question”When the agent needs information from you, Chat shows a Question card.
Depending on the question, you may see:
- one selectable choice;
- multiple choices;
- a free-form answer field.
Choose or type the answer and press Send answer.
Your answer resumes that Agent Runtime run.
5. Understand why a prompt appeared
Section titled “5. Understand why a prompt appeared”Your Agent profile’s Approval mode influences normal tool behavior:
- Withhold risky tools
- Pause for approval
- Auto-run except high risk
- Auto-run all allowed tools
Per-tool overrides and argument-sensitive safeguards can make a specific call stricter than the general profile setting.
Current restart boundary
Section titled “Current restart boundary”Approvals and questions can pause and resume normally while the backend process stays running.
Restart-safe continuation for longer work and deeper delegation is part of the separate Cognitive runtime Preview. If an important approval or question is still waiting, finish it before intentionally restarting the backend.
Expected result
Section titled “Expected result”After approvals/questions are resolved, the run continues through Agent Runtime and Chat records the resulting tool/runtime trace and assistant output.