# Give the agent an ongoing responsibility

Extend the person's agent environment so it can carry a useful responsibility across successive runs. Build it extremely well. The person should be able to explain what matters, give the agent appropriate authority, and return to work that has advanced with that understanding intact. Follow the [shared launch](../shared/launch.md).

Start from an existing responsibility, such as the [overnight researcher](../background-services/overnight-researcher.md), or choose another recurring job the person wants handled. Reuse the existing coding agent, account, tools, and scheduling capabilities.

## Define the responsibility through its purpose

Learn what makes the work valuable and when it deserves attention. A researcher might follow developments relevant to a practical question, pursue a promising lead over several nights, and report when the evidence changes its explanation.

Preserve the purpose, available sources and tools, chosen resource budget, and where results belong. Make the agent's scope of action clear: what it can carry out independently and which decisions it should bring back to the person.

Give the agent room for judgment within that responsibility. It can choose what is worth investigating, recognize that nothing useful needs doing today, or explain why an unresolved question deserves another focused packet. Let the actual work determine the report's length.

## Carry understanding between runs

Make current direction, previous results, unfinished questions, and useful observations available when the agent resumes. Use the [project notebook](project-notebook.md) or appropriate existing records, preserving the actual sources of current state.

Recognize work already completed and continue from it. Keep new instructions and changed priorities apparent so a later run serves the person's current purpose. Preserve valuable earlier work when a question is paused or replaced.

Give substantial stages their own attention through the shared autonomy-loop method. Finish a meaningful packet, assess what it established, and select the next one while authorized work remains. A saved draft can receive its editing pass in a later invocation; finishing the draft alone does not finish that larger assignment.

Teach the distinction between this working method and provider goal tracking, including OpenAI's goal feature. Focused packets and honest checkpoints organize the responsibility; the installed harness supplies whatever scheduling and continuation actually occur.

## Make the wake and controls real

Establish a supported schedule, event trigger, or continuation mechanism for the person's environment. Bring the right responsibility, tools, and relevant skills into the actual resumed run. Explain which machine or service must remain available and what happens after downtime.

Show whether a run started, what it worked on, what it finished, and what is waiting or blocked. Keep a future scheduled run distinct from work currently happening. Let the person inspect the responsibility, run it now, revise it, pause it, or end it.

Use the agreed budget and notification preferences. Report a problem when it affects the responsibility, preserve completed work, and make a useful return to the task possible after the cause is resolved.

## Deliver a responsibility that can continue

Complete the instructions, context, tool access, scheduling, and controls. Exercise the real launch path, carry a useful piece of work across a later run, and show how the person changes the direction. Include the ordinary outcome where the agent has nothing worthwhile to add.

Judge whether the continuing work serves the person's purpose. Use their response to improve the agent's understanding of the responsibility, preserving what already works.

Leave its records, results, and setup recoverable through the shared launch's ownership guidance. The learner should understand what wakes the agent, what gives it context, and what lets it make good decisions when they are doing something else.