Atlarix 15.11 shipped with a feature that solves one of the most annoying problems in AI-assisted development: self-review bias. The new cross-session review capability allows one Atlarix session to review another's pending work before it is applied to the codebase. This isn't just a read-only observation mode; the reviewing session can actively approve, deny, or replace the proposed changes. For builders who have ever watched an AI confidently write buggy code because it was too close to the problem, this adds a necessary layer of separation.

Breaking the Echo Chamber

The core argument for this feature is simple: a session that wrote the code is the worst possible reviewer of that code. It shares the same context window, the same assumptions, and the same blind spots that led to the initial output. By introducing a second session—whether started fresh or forked earlier—developers can see the change without the baggage of the creation process. The reviewer session asks the questions the original session assumed were already answered, catching logical errors and edge cases that often slip through single-pass generation.

Implementation and Workflow

Setting up this workflow requires starting two sessions in the same workspace. In the first session, you let the AI build a feature until it reaches a plan or a permission request. In the second session, you run list_sessions, which reveals the first session with hasApproval: true or a plan awaiting action. You can then attach the pending state using send_message with review: true, or answer directly using approval: or plan: approve. The target session visibly acknowledges who answered, ensuring the review process is transparent rather than silent.

Safety Limits for High-Risk Actions

Not all changes are created equal, and Atlarix draws a hard line at the request level. High-risk requests—those capable of causing data loss or security issues—cannot be approved by a peer session. These critical operations still require the human owner's direct input. This design choice reinforces that while AI sessions can act as contributors and reviewers, they are not administrators with full veto power over destructive actions. It’s a pragmatic balance between automation speed and safety.

Key Takeaways

  • Cross-session review allows one Atlarix session to approve, deny, or replace another's pending plans and edits.
  • The feature mitigates self-review bias by using a second session with fresh context to catch assumptions.
  • High-risk operations remain locked behind human approval, preventing peer sessions from triggering data loss.
  • Reviewers can view full plans via list_sessions with plan_of and respond via send_message.

The Bottom Line

Self-review is a failure mode we should have designed out of AI agents years ago. Atlarix 15.11 finally treats AI sessions as distinct actors with separate contexts, which is the only way to actually catch hallucinations before they hit production.