Using the AI Agent — One-off Mode
One-off mode runs the agent autonomously on a single prompt, without any interaction. You send a task, the agent works through it to completion, and the output streams back to your browser.
Use one-off mode for:
- Well-defined, self-contained tasks with a clear expected outcome
- Tasks where you don't need to answer questions mid-run
- Repetitive operations that follow a predictable pattern
How it differs from interactive mode
| Interactive | One-off | |
|---|---|---|
| Session type | Live, persistent | Single run |
| Mid-run questions | Yes — you answer in the browser | No — the agent decides automatically |
| Best for | Complex, exploratory work | Specific, well-scoped tasks |
| Reconnection | Yes (5-minute window) | No |
Writing effective one-off prompts
Since the agent can't ask you questions, your prompt needs to be complete enough that the agent can make all decisions on its own.
Less effective:
Update the RBAC config
More effective:
In
src/config/rbac.py, add a new role calledVisitorwith the following permissions: read access to zones A, B, and C; no write access; no admin access. Use the same structure as the existingContractorrole. Do not modify any other roles.
Include:
- The specific files or areas to work on
- The exact change to make
- Any constraints or things to leave unchanged
- The expected outcome
Thread continuity
When using DeepAgents in one-off mode, the platform remembers the conversation thread between prompts. This means you can send a follow-up prompt and the agent will have context from the previous run.
Claude Code one-off mode does not retain context between separate runs — each prompt starts fresh.
After the run
One-off runs don't automatically commit changes. After the agent finishes:
- Review the changes in the file sidebar
- Check git diff to confirm the changes are correct
- Commit and push when ready (see Git Operations)