ARSI
The agent should be sharper next month than it is today.
Acumane Recursive Self-Improvement is a force multiplier for the agents you already run. It does not replace the method. It measures a round of work, keeps what held up, and offers a version your team can approve.
Most assistants forget the correction you typed last Tuesday. ARSI writes that correction into the agent as a version, with a changelog, and the same publish path you use for a manual edit. Client files stay in their workspace. The improvement moves. If the new version misses the suite, it stays a proposal.
Approval is the default
Every agent starts in suggest mode. You see the pending instruction, skill, example, and memory changes, the suite score before and after, and you publish or dismiss each one. Auto-apply is a switch for teams who want the gate to be the only stop. Off means the loop does not run.
The dashboard charts the suite score of the instructions and of the selected candidate, round by round, so improvement is a number you can point at rather than a feeling about the last chat.
What this release runs without a model key
With a provider key, Predict calls a hosted judge through the AI SDK. Without one, the same step uses a mock outcome model that still ranks candidates and still refuses a regression. The loop, the versions, and the approval screen do not wait on that key.
The cycle follows the research framework RSI-Anything (Apache-2.0). Acumane does not ship that codebase. The paper is cited in the product notes.