What happens when you reject an AI agent's work in Bees

When you reject an AI agent's work in Bees, the process goes beyond simply starting over or editing the prompt. Bees implements a middle ground in the form of a reviewer agent. This reviewer checks the work against the goal and the evidence recorded by Bees, without considering the worker's private reasoning. If the work fails to meet the requirements, the reviewer sends the task back to the work stage with detailed feedback, starting a new attempt.

This process is capped with a limit on retries before the goal transitions to a "Needs your attention" state.

In addition, Bees offers the option for users to directly review certain stages or goals by providing "Approve" or "Reject" choices. When rejecting, a reason is required. This reason is scoped, meaning its impact depends on the context: for a single goal, the feedback is given back to the agent for revision, while for scheduled jobs, the reason is added to a playbook for that specific job.

This playbook is managed by a specialist, a specific agent tailored for that job. When an agent starts handling a named schedule for the first time, a specialist is created, inheriting the base agent but adding job-specific notes. This ensures that changes are isolated to the relevant job and do not affect other aspects of the agent's behavior.

Notably, one-off goals do not influence future behavior, meaning there's no automatic learning based on rejection scores. The reviewer's decision doesn't alter the base agent for the entire organization, nor does it change the agent's base functionality due to a rejection. This distinction between denying an agent's tool action and rejecting the work itself is maintained. For instance, rejecting an email sending action is a permission decision for the moment and doesn't lead to a permanent rule.

Bees provides transparency into the review process. Each specialist and playbook can be inspected under Process Runs → Schedules, showing the playbook, its revision number, and the history of changes. These changes can be edited, undone, or reset to the base agent state. This approach prioritizes human oversight over automatic learning, recognizing that a poorly reasoned rule based on opaque criteria can be more detrimental than no rule at all.

Bees is open source, available under MIT or Apache 2.0 licenses, and supports various AI models like Codex, Claude Code, or local models. It is accessible on Mac, Windows, and Linux platforms at https://bees.bot/download/ and its source code is hosted at https://github.com/Bees-bot/bees-desktop.