How to Hand Off AI Agent Work When You Hit Your weekly Usage Limit

When working with AI agents, hitting the weekly usage limit can be a frustrating challenge. Many developers face this problem daily, and the usual solutions of waiting for a reset or sharing credentials are not ideal. Waiting for a reset may not be practical, as the context built up during the session may be lost after some time. Sharing credentials, on the other hand, creates security risks as it involves exposing sensitive information without proper auditing or expiration.

A third option is to focus on documenting the task properly, rather than just sharing credentials. The issue with handing off AI work lies not in the credentials, but in the lack of a clear understanding of what the task is and its specifications. Most handoffs fail even with perfect credentials because the person taking over cannot reconstruct the task's goal, constraints, and definition of done.

To avoid this, tasks should be described using three key elements:

1. The goal, in outcome terms: Instead of stating an activity like "refactor the auth module," the goal should focus on the desired outcome, such as "Auth module should stop making a network call per request and target one call per session."

2. The constraints: It is essential to document any non-obvious limitations, such as "Don't touch X," "We tried Y and it deadlocked," or "The tests in Z are flaky, so ignore them." These constraints may not be obvious and can cause issues if not addressed.

3. The definition of done: A clear definition of done should be specified, including the exact command to run and the expected output. This helps ensure that the task is completed successfully and to the required standards.

A simple template can be used to fill in these details:

## Goal

Outcome, stated so it can be checked

## Constraints

- Thing that must not change, and why

- Approach already tried and rejected, and why

## Done when

Command to run, and the expected output

## Context

Links: the branch, the issue, the failing CI run

Writing the task description in this format takes only a few seconds but can save an hour of back-and-forth communication.

The real issue lies in sharing the account, rather than the task. When handing off AI work, the unit that should be shared is the task itself, not the account. Each person should authenticate as themselves, using their own subscription and a scoped, revocable token. This approach ensures that the work is isolated, and there is no shared secret with no audit trail, scope, or expiry.

Using scoped and revocable tokens instead of a shared key helps prevent unauthorized access and limits the impact if a token is compromised. Additionally, tokens should expire on their own, further enhancing security.

Wagglet is a tool built around this handoff format, allowing teammates to document the task and another person to claim it and run it on their own subscription with scoped credentials. The tool also provides a purpose-built format and security model for sharing tasks.