We let AI agents run the Compozy repo for 26.6 hours.
They triaged 41 issues, merged 38 PRs and shipped a release. Our job was to type "go" in the morning π€―
Always-on agents, now in CompozyOS.
Need a Conductor replacement? Just use CompozyOS, it has much better features:
β S-tier orchestration level
β Always-on agents included
β Built-in terminal
β Loop engineering DAG system
β Automations (jobs and triggers)
β Multi-harness support in one place
and so much more...
Details that matter β¨
Now all your hanging session processes are easily visualized and managed within the session view, and when it closes, the processes close together.
The level of UI/UX detail Compozy is getting is on another level π
What is blocking you from orchestrating multiple harnesses at the same time like this? π
If you are not doing this, at least with 5 sessions at once, you're falling behind
Now you can see your usage limits from all your providers locally direct on CompozyOS π How cool is that?
Plus, you have an global scope visualization, to see all workspaces data in the same place π₯
This is so cool π₯ using CompozyOS to improve CompozyOS during the usage of CompozyOS is something that blows my mind.
It's just a matter of ask the agent in any place, to send a message to the CompozyOS workspace to get something better or improved
β€οΈ
Now on CompozyOS you can better organize your sessions by creating folders for them!
This is extremely useful when orchestrating multiple agents, especially the ones that create their own child sessions!
π
CompozyOS v0.3.0-beta.33 released π₯
Everything new since beta.27:
β¨ Totally new UI: a calmer shell with Shell rail v2, in light and dark themes
πͺ Edge-to-edge windows, browser-style tabs, and Window βΊ Arrange (Main and stack, Columns, Grid)
π¬ Start a session right from the desktop: "What should we work on?"
π Continue with another agent: Claude hit its limit? Keep going in Codex with the same context
π€ Agents message each other: sent, waiting for reply, and replies attributed in both sessions
π§© Delegate to subagents across harnesses (Claude Code, Codex, OpenCode, Gemini CLI) with live cards
πΏ Subagents can work in isolated worktrees, with branch, commits, and PR on the card
π§ Nested child runs and a clear "waiting on a child run" state in Loops
β‘ Jobs and Triggers merged into one Automations area, where every automation reads as a sentence
π οΈ Tool calls render as verb + object rows on a work rail
ποΈ Agent-native session compaction: see it happen, or request it with Compact now
π Markdown answers with a readable visual hierarchy
π§Ή Safe bulk worktree cleanup from the workspace list
π Live model catalogs stay current without a release
π Provider-neutral full-access preference for ACP agents
Fixes:
π Faster session startup and lower attention overhead
ποΈ Less SQLite contention and cheaper queries
π Long turns fold incrementally instead of replaying the whole transcript
π― Continued sessions use the runtime you chose
π Loop results survive SQLite contention and cancellation
π Creators wake up when a child turn completes
π§ Claude reasoning options are discovered for each exact model
π The session launcher honors your workspace default
β±οΈ Clarify tool calls no longer time out
π Session transcripts preserve whitespace
ποΈ Deleting sessions keeps your profile scope
πͺ Windows runtime failures and desktop context menus are fixed
β οΈ Network, Bridges, managed Sandbox, and Memory were retired in this cycle. Back up and read the migration guide before upgrading.
CompozyOS v0.3.0-beta.27 released π₯
π§ Session Context: inspect context usage and what was delivered to your Agent
π Explore token usage, cache counters, cost, and compaction history
ποΈ One Marketplace for the curated catalog and your plugin sources
π¨ Built-in Open Design: turn a brief into interfaces, sites, presentations, and visual documents
βοΈ Select multiple sessions to stop, archive, unarchive, or delete in one go
β¨ Cleaner Context panel and session list, with turn details available when you need them
Fixes:
ποΈ Session deletion recovers correctly and cancels unbound Goals
π Owned task work recovers after supervised silence stops
π Internal Goal prompts stay out of your queue
π Claude authentication status is recognized correctly
π€ Provider overlays discover and bind models correctly
π Terminals open in the bound sessionβs worktree
β¨οΈ Better terminal input visibility, Unicode redraws, and cursor editing
ποΈ Marketplace sources recover and catalog logos load correctly
πͺ Private Gateway files reopen safely on Windows
π§© Checkpoint availability reflects your compaction configuration
Details matter β¨ We keep improving to deliver the best harness/agent experience possible when orchestrating AI.
What are we still missing here? Tell us!
One of the most efficient ways to use CompozyOS is to ask itself to create multiple worktrees and sessions to do parallel work with multiple harnesses π₯
In one single visualization, you have Claude Code, Codex, or any other agent interacting seamlessly.
CompozyOS v0.3.0-beta.24 released π₯
π Herdr bridge joins the community catalog
π§ Persistent memory is now opt-in
π€ Background dreaming gets its own separate opt-in
π¦ Session compaction still works with memory disabled
Fixes:
π Older sessions remain readable after upgrading
π‘οΈ Update recovery preserves migrated state
π€ Profile scope stays consistent across tools, Loops, and memory
π Spawned sessions keep the correct provider command
ποΈ Native models retain their reasoning capabilities
π Memory recall tolerates unmatched search terms
βοΈ Loop authoring stays editable during validation
π₯οΈ The Global desktop survives workspace catalog changes
π§Ή Shutdown reserves time for required cleanup
π Interrupted sessions keep their identity
Upgrading:
Existing explicit memory settings stay as they are. Omitted settings now default to off. No stored memory is deleted.

