PRIVATE PREVIEW · VERSION 2026-10-11

Usage & AI limitations

Powerful helpers. Human responsibility.

Effective and last updated: 11 October 2026

AI review is not a guarantee

Both the implementing agent and reviewer can miss defects, invent explanations or misread requirements. Tests may be incomplete, use mocks or fail to run. “Implemented,” “verified,” and “verification blocked” mean different things. Check what actually ran, which revision was checked and which scenarios remain unverified. Screenshots show the state captured at that time, not every possible user journey.

Agents can change real files and run commands

A delegated task can edit source code, install dependencies and execute development commands under the host’s permissions. Separate checkouts help isolate repository changes; they do not automatically isolate credentials, ports, databases or network services. Avoid production credentials and use appropriate development environments.

With PR delivery selected at approval, Relay commits and pushes the reviewed task branches and opens or updates PRs using gh on the execution laptop. Existing local-only tasks are not published automatically. Agents are instructed not to push themselves; Relay’s publishing step does not merge or deploy. These instructions are not a universal operating-system or network security boundary. Maintain provider permissions, repository protections and release controls independently.

Feedback helps; the owner decides

Participants do not have to respond before the owner accepts a verified result. Existing feedback and outstanding change requests stay visible in the review record. The owner is responsible for weighing that input. Acceptance records a workflow decision and does not itself merge a pull request, release software or override branch protections.

Your accounts and charges

Local Codex and Claude Code runs use the execution host’s configured accounts and limits. Managed runs may incur separate API charges. Usage displays can be stale or unavailable; provider billing is authoritative. Repair loops, tests and review may consume additional time and allowance even if the task is not completed.

Shared and offline behavior

Cloud discussion is available independently of an execution laptop. Local agent work needs the host and required services to remain available. A disconnected host is not automatically replaced, and closing an app is not a reliable way to cancel every local or managed process. Use the pause control and check the reported status.

Progress is an observation, not a guaranteed percentage complete. A stopped or disconnected run may need recovery before work continues. Keep your own repository backups; Relay is not a backup or disaster-recovery service.

Private preview

Features and availability can change. The current Mac preview targets Apple silicon and is development-signed, not notarized for general distribution. Use the Figureit-provided installer and follow your organization’s device policies.

Relay coordinates third-party tools; it is not a security audit, compliance certification or substitute for qualified review where needed. Provider terms and usage policies still apply. See Terms of use and Privacy notice, or contact admin@figureit.ai.