What shipped.

1. v0.60

   Sep 26, 2026

   Quiet when you're away

   • With the window hidden and nothing running, the app closes idle browsers and slows its background work to once every 5 minutes.
   • Your answers on a question card now reach the agent that asked, in its next session.
   • An open conversation no longer pauses the schedule forever: it hands the work back after 60 idle minutes.
   • Session costs read correctly. They had been counting the conversation's running total, about 9 times too high.
   • The Run card shows what a Run is waiting for you to answer.

2. v0.59.3

   Sep 24

   Background conversations cost nothing

   • Conversations you aren't looking at stop rendering: the page's memory dropped from about 600 MB to under 200 MB with six open.
   • A turn that hangs ends on its own and is recorded as interrupted, instead of blocking the agent.

3. v0.59.2

   Sep 23

   Each turn thinks as hard as its job needs

   • Building and fixing get deep thinking, reviews get high, sorting and triage get light. The reply shows which it used.
   • Say “think hard” or share a plan and live mode thinks deeper for that message.

4. v0.59

   Sep 22

   Opus 5.5 by default

   • Opus 5.5 is the default model once your Claude Code can run it (2.1.280 or later).
   • “Left running by agents” lists servers and browsers that outlived their turn, with a Stop button.
   • Fixed turns dying with “low max file descriptors” on busy machines.

5. v0.53 – 0.58

   Sep 19 – 21

   Everything the app runs, in one list

   • Settings shows a process census: every browser, server and agent turn the app started, and how much memory each uses.
   • Processes left behind by a crash or a force quit are cleaned up at the next launch.

6. v0.49 – 0.52

   Sep 17 – 18

   A new design

   • A new mark and icon, warm light and dark themes from one palette, and a sidebar that folds into a rail.
   • Live mode reads as one white page; the composer is its only card.
   • One “Needs you” list per workspace: decisions, drafted replies, Run gates and failed runs.

7. v0.33 – 0.48

   Sep 9 – 14

   Your inbox becomes the agents' work

   • Slack threads and meeting commitments become signals, routed to the right workspace with the evidence for the choice.
   • Replies are drafted in the client's own thread and wait for your OK.
   • A call transcript becomes a proposed plan; a screenshot dropped on the card becomes a signal.
   • A taste judge reviews screens before and after a change, learning from your comments.
   • Approve plans and routes from your phone on Telegram.
   • Independent waves of a Run build in parallel.

8. v0.31 – 0.32

   Sep 2 – 6

   Runs

   • Hand the app a plan in waves. It works through it one session at a time, and a wave closes only when an evaluator agrees the result was hit.
   • The model picker hides models your installed Claude Code can't run.

9. v0.29 – 0.30

   Aug 28 – Sep 1

   Conversations you can come back to

   • Past conversations reopen with one click, can be named, and show when they have something unread.
   • Each workspace shows when its next run is, with Pause and Resume.
   • New models, like Fable 5.1, arrive without an app update.
   • Every file an agent writes is listed in the workspace library.

10. v0.27 – 0.28

    Jul 27 – Aug 14

    iOS apps

    • Agents create, build and run iOS apps in the Simulator, and check them with flows that tap, type and assert.
    • Watch the Simulator live and take over when needed; ship builds to TestFlight.
    • Workspace memory rotates on its own, so it never grows past what a session can read.

11. v0.24 – 0.26

    Jul 23 – 26

    Opus 5 and quiet updates

    • Claude Opus 5 support, and it became the default.
    • Updates download in the background and apply the next time you open the app.
    • A conversation survives hitting a usage limit and picks up where it stopped.

12. v0.22 – 0.23

    Jul 17 – 21

    The browser

    • Each agent gets its own browser. Log in once and it works from your sessions.
    • Watch it live beside the conversation, and take over for a login or 2FA code.
    • Every page it visits is logged, and sessions can be saved as replays.
    • Action gating, rate limits and quiet hours; an overnight QA patrol with a morning digest.

Earlier releases (March to July) built the base: the agent loop, memory, the wiki, Telegram and email.