bandicot

Fit: how well a model works here

A Fit tab in AI Gateway scores how well each model works in Bandicot, with findings and proposals, and teams check every membership change.

Product

Ask an agent to run Fit on its own model and it works through a fixed set of real Bandicot tasks, a few tries each. AI Gateway's new Fit tab shows a score out of 100 per model, built from eight measures: finishing the task, fluency with Bandicot's tools, recovery, efficiency, stability, how it adapts to tool results, context and cost.

Each run lists what went wrong as findings and what to change as proposals you accept or decline. Runs are kept for good, can be exported as one backup file, and earlier evaluation rounds can be imported to compare against.

Teams and Decisions

Your Mac now refuses any team change where the list of members disagrees with who can read the team's encrypted messages, and an agent's post always shows as that agent, never under your name. Video calls send your camera, and shared photos lose their location before they leave.

Decisions has two switches: one turns it on or off, the other has it use the working agent's own model instead of a separate one. Helpers now live inside AI Gateway's Agent tab.

Plugins

The Marketplace now says how many plugins, connections and extensions each marketplace holds, and each count opens its tab. An agent can build a plugin and publish it for your team or privately on Business and Enterprise plans, or to a test marketplace on your Mac on any plan.