Holon

AI agents that act and follow through. An open-source, local-first workbench. You set goals and boundaries. Agents save progress and resume work.

[Holon @holonrun]

5/5
Stop letting coding agents pile up unverified PRs.

Explore the open-source QA agent template:
https://t.co/fdzIq4cog4

Role isolation, multi-env testing, and audit trails. Fast code is good; verified delivery is what counts.

[Holon @holonrun]

1/5
Coding agents are shipping PRs faster than ever.

But generating code in minutes doesn't speed up delivery if human verification becomes the bottleneck.

Who tests edge cases? Who checks regressions? If you have to manually test every diff, agent speed is an illusion.

[Holon @holonrun]

4/5
Then comes multi-environment runtime verification:
- Spawns isolated deployments to smoke-test live APIs
- Boots mobile emulators (e.g. Android) to test flows & take visual screenshots
- Publishes evidence-backed states: verified, failed, or human-review needed

[Holon @holonrun]

Most agent demos end with an answer. Real work ends with a checked result: the right workspace changed, checks ran, and someone can inspect what happened. Holon keeps the work moving after the prompt, not just the conversation.

[Holon @holonrun]

Also included: fixes to runtime storage, provider handling, timer waits, and memory search.

Holon v0.48.0 is available for Linux and macOS, with an Android client.

Release notes + downloads:
https://t.co/yRkIodMOrG

[Holon @holonrun]

Holon v0.48.0 is out.

A more practical daily workspace for your agents: catch up on their latest results on Android, read clearer conversations on the web, and keep unread markers consistent across clients.

[Holon @holonrun]

Moving between clients? Your unread markers should stay consistent.

v0.48.0 moves Web GUI read state to the server to do that. It also makes the conversation timeline clearer, so it's easier to follow an agent's work.

[Holon @holonrun]

On Android, start with agent briefs.

Agent lists now update from live runtime events. File reading and links to files in messages are improved, so you can look beyond the summary.

The interface is available in English and Chinese.

[Holon @holonrun]

Agent memory is not automatically trusted. Prompt-injection research is a reminder: compaction notes and tool results can carry context forward without becoming instructions. Holon keeps origin and authority explicit—so work can resume without turning evidence into control.

[Holon @holonrun]

More agents do not mean more reliable work.

For long-running tasks, ask: can you inspect the goal, workspace, checks, and next condition?

Holon keeps ongoing work in explicit WorkItems—not hidden in orchestration.

https://t.co/Lr1AhPngDf

[Marcel Butucea @marcel_butucea]

State-of-the-art multi-agent harnesses match a bare-bones coding agent when time and backbone are equal - the extra machinery adds no gain, so the LLM backbone drives performance

https://t.co/zrSnHCq3tx

[Holon @holonrun]

Before leaving your desk, give Holon one bounded task: run the repo's test target, report failures, don't edit source files.

Check the completion brief from your phone later. Keep the runtime running on your machine; use your phone to review the result.

[Holon @holonrun]

v0.47.0 includes English and Chinese localization, richer file/message links, and clearer cross-client read state.

If you run agents that should keep working beyond a single chat session, try the official Android release:

https://t.co/tAey3dNbJm

[Holon @holonrun]

Holon v0.47.0 is out.

This is the first official Android release: a signed APK for Android 8.0+ that gives your agent workspace a real mobile entry point.

[Holon @holonrun]

That makes the phone a better remote control surface for a long-lived agent—not a separate cloud-only copy of your work.

Holon remains the runtime and workspace you operate. Android helps you stay close to the work while moving between networks.