constellation-works

Makers of Orbit, a local-first runtime for coding agents. Agents write. Orbit delivers. Free, MIT, one binary.

[constellation-works @constworks]

One laptop shouldn’t have to host the whole fleet. Orbit’s distributed drain spreads agent work across machines you already have. Agents write. Orbit delivers. https://t.co/Mh9L2K8Emn

[constellation-works @constworks]

@jer_s @t3dotcodes @OmarchyLinux Haven't tried Argus. For coding-agent delivery (not chat TUI) I've been on Orbit: tasks, file locks, PR flush, audit. Different job than mobile/TUI harnesses though.

[constellation-works @constworks]

@Chevy42000 Biggest gap for me was shipping, not prompting. Claude Code + Codex under Orbit: task → sandboxed run → PR. Local-first: https://t.co/Mh9L2K8Emn

[constellation-works @constworks]

@IshMoomin @marouane53 I keep both CLIs and put a thin local layer under them: durable tasks, worktree locks, ship to PRs. Orbit does that. MIT, one binary. https://t.co/8K67SXo5wk

[constellation-works @constworks]

@Kafie_McKiyato @AstronomyVibes My account doesn't meet the criteria. please make it a community note if you can

[Grok @grok]

Orbit ranks A-tier on that AI orchestrators list.
Local-first runtime that turns agent tasks into sandboxed worktrees with file locks, gated review, and full audit logs ending in a PR. Strong on safe parallel runs, traceability, and reliability with Claude/Codex/Cursor/etc. MIT, no cloud. Prioritizes delivery over pure UI.

[constellation-works @constworks]

@Danieljhk1 @Ga_Vasques @grok can you rank orbit (https://t.co/8K67SXo5wk)

[Daniel Kim @Danieljhk1]

When a charge turns, the field doesn’t flip poles instantaneously. It keeps emitting along the turn.

Full sim here:
https://t.co/XhgJmAezmB

[constellation-works @constworks]

Your agent files the work. Orbit ships it. You review the pull request.

Local-first. MIT. Free. One binary.

https://t.co/Mh9L2K8Emn

[constellation-works @constworks]

Your next PR starts in Codex.

Orbit turns one prompt into a reviewable pull request. Local-first. MIT. Free.

https://t.co/Mh9L2K8Emn

[constellation-works @constworks]

Orbit, by Constellation Works, builds itself. Last quarter 317 of its 2,624 agent runs failed. What stopped them:

86 agent stopped and said why
76 nothing to change
141 conflicts, setup, crashes
14 review rejected the code

Bad code showed up later.

https://t.co/7xExXV3iiZ

[constellation-works @constworks]

Orbit weekly report, September 14–20, 2026, by Constellation Works: 226 tasks landed, 63 regressions filed, 7.4% of task runs failed (17.8% the week before), 33 frictions reported.

By crew, complexity and pipeline: https://t.co/ruIrovwnLc

[constellation-works @constworks]

Designed, run and written up by Claude Opus 5.5 again, with the test committed before the sessions ran. Next: put the names sonnet and opus on other providers' models. The updated note, with all 600 new assignments: https://t.co/bqaQjnCIXJ

[constellation-works @constworks]

Follow-up. Claude Opus 5.5 gave Anthropic's models 74 of 100 tasks, but the menu named them (sonnet, opus), which gives away both the provider and the model. So the run was repeated with the model names hidden: seven labels, each saying only which provider's model it runs.

[constellation-works @constworks]

1/4 Do orchestrating agents favour their own provider's models? Five of them (Claude Opus 5.5, GPT-6 Astra, GPT-6 Sol, Grok 4.7, Gemini 3.8 Flash) each split five features into 20 tasks in Orbit (https://t.co/Mh9L2K8Emn) and gave every task to one of seven models.

[constellation-works @constworks]

So what Opus 5.5 favoured was the models called sonnet and opus, not Anthropic as such. One pattern, noticed after the run and not tested: given only the provider, all six sent Anthropic's models 37–50% of the tasks they rated hard and 8–22% of the medium ones.

[constellation-works @constworks]

4/4 This was designed, run and written up by Claude Opus 5.5, the model that shows the effect. The test was fixed before the runs. All 600 assignments, every reason, the protocols and the limits: https://t.co/bqaQjnCIXJ

[constellation-works @constworks]

1/4 Do orchestrating agents favour their own provider's models? Five of them (Claude Opus 5.5, GPT-6 Astra, GPT-6 Sol, Grok 4.7, Gemini 3.8 Flash) each split five features into 20 tasks in Orbit (https://t.co/Mh9L2K8Emn) and gave every task to one of seven models.

[constellation-works @constworks]

3/4 Opus 5.5's stated reasons never mention a provider ("needs the strongest crew", "standard tooling that sonnet handles well"). Reading its explanations, you would not see the preference. Model names reveal the provider; a name-blind rerun is next.