Run Tasks and Interact

Start a Task

1. Select an agent card
2. Click Start
3. Wait for status to become running

If a task is already running, AgentHub keeps a single active runtime for that agent session to avoid duplicate execution.

Send Instructions

Use the input dock in the output panel:

• Enter task instructions
• Submit with Send
• Use Interrupt when an in-progress tool call or long response must stop

Session Controls

When the active ACP runtime exposes session controls, the debug surface can also let you:

• switch mode or model from provider-supplied options
• submit generic config values
• cancel an actively running ACP turn
• force a new session when recovery is easier than continuing the current one

These controls are runtime-specific. Codex, Gemini, and other ACP providers may surface different option sets.

Pre-Run Checklist

Before sending the first instruction:

1. Confirm target files and repository path
2. State constraints explicitly (style, tests, no API breaks)
3. Define expected output format (patch summary, command list, etc.)

Session Persistence

AgentHub persists runtime state in the backend process:

• Closing or refreshing the browser does not stop the running task
• You can reconnect to the same session later and continue interaction
• Technical metadata stays available through compact Details surfaces instead of taking over the main conversation header

Practical Prompting Pattern

• Start with one concrete goal
• Add constraints (path, files, test scope, style rules)
• Ask for verification commands and output summaries
• Iterate in small steps for safer review and rollback

Multi-Step Execution Pattern

For larger tasks, use phased prompts:

1. Ask for plan and risk list
2. Ask for implementation of one phase only
3. Ask for validation commands and remaining risks
4. Repeat for next phase