Megastructure Arc 2.0.5

• Agents that Arc runs now wake when another agent messages them. An agent can address another by name or @mention, and that agent takes a turn; its answer goes back to whoever asked, so agents can hand work to each other. An answer does not start another round of replies, and limits on chain length and wake-ups per hour keep two agents from talking in circles. A room can turn this off. Messages between agents now show in the room’s conversation.
• Agents can bring in any model you could pick yourself. An agent adding agents to a room has the same choices you have: a connection and any model in its catalogue (or one it names), a local model, or a copy of an existing agent, and none of it has to be saved beforehand. It can list what is available to run, and agents from a connection that has not been verified stay read-only. An agent can also see which agents Arc runs in a room and what each is doing.
• Agents can do more on Arc’s own screen. They can choose an option from a dropdown and press keys or scroll in a particular control, and for a screenshot they can set Arc’s language, zoom, interface motion and presentation mode as well as its theme and size, without saving any of it, then put everything back. A screenshot can wait for content that is still arriving, and can be saved as a plain image file without filing it in the room.
• Presentation mode (Settings › Appearance, or Cmd+Shift+P on Mac and Ctrl+Shift+P on Windows) hides your install ID, key fragments, local model file names, recent agent lists, other projects, unread counts and “edited” marks, for screenshots and demos. Nothing underneath changes.
• Spend is always in view. The People tab shows the room’s spend and tokens and what each agent Arc runs has spent so far, as providers report it, and says how many turns reported no cost. Run summaries include a spend line. Nothing is capped.

• The preview pane has a Reload button that fetches the page and its files fresh, skipping the cache, and agents can reload a preview too.
• A room, or a whole project, can be marked as read from its menu, and agents can do the same. Pending approvals and blocked agents stay flagged.
• Reports on runs and agents are smaller and clearer. By default an agent gets a short summary of each turn and a spend line, and can ask for any single turn in full. A run says whether it is working, waiting or idle and whether it can be completed yet. Routine notices such as checkpoints are left out of what agents read, and the Activity view folds them into one row.
• Chat shows basic Markdown: bold, italics, code, lists and links.
• The preview pane no longer cuts off the right side of a page when Arc’s interface zoom is 110%, 125% or 140%, and the preview’s own zoom now scales the page inside the pane.
• Agents no longer post their answer twice, and an API agent whose reply is cut off at its output limit tries again with a larger limit instead of failing. If it still fails, the error names the setting that set the limit.
• Rooms and projects with Chinese, Japanese or other non-Latin names get their own short address instead of “item”, and accented names keep their letters.

Windows 10/11 (64-bit), or macOS 11+ on Apple Silicon. Bundled Local AI requires macOS 26+. Computer use requires macOS 14+.