Notes

Research and operating data from building and running Orbit, a local-first runtime for coding agents. Each note publishes the data behind it.

• November 3, 2026

  Crew rankings by tier: Haiku holds its own on medium work

  How 1,432 task attempts by coding agents in Orbit's own repository from October 27 to November 3, 2026 ranked, by crew and complexity tier. On medium work, Grok, Muse Wonder and Sol led, and Haiku, a small model outside the medium pool, ranked fourth of eight, ahead of Sonnet. Opus led hard work on the score, but had 30 more escaped regressions than expected. Nearly every crew landed nearly every task. Rework separated them, and escaped regressions, which the ranking leaves out, change the picture for some crews.

• October 31, 2026

  Cheaper models on medium work: half the cost, no verdict

  For three days, half of Orbit's medium-complexity tasks went to cheaper models (Haiku, Luna, Muse Spark). Where cost could be priced, those tasks cost about half as much, with flat recoveries and review repairs. The experiment still ended inconclusive. Delivery dipped 2.6 points, the control arm's models changed mid-run, and the first cost readout pointed the wrong way.

• October 2, 2026

  Most failed agent runs weren't about the code

  Over a quarter of Orbit building itself, 317 of 2,624 agent delivery runs failed. A review step rejected the change in 14 of them. Agents stopped and gave a reason in 86 and found nothing to change in 76; the rest stopped on concurrent work, the machine or the agent process. The code's problems showed up after success: at least 186 delivered tasks later drew a regression report.

• September 27, 2026

  Orbit weekly report: September 14–20, 2026

  226 tasks landed, 63 regressions were filed, 7.4% of task runs failed (17.8% the week before) and 33 frictions were reported. Regressions by crew, completion and implementation time by complexity, and pipeline reliability from Constellation Works' own Orbit store.

• September 26, 2026

  Do orchestrating agents pick their own provider's models?

  Five orchestrating agents split five features into 500 Orbit tasks and chose a model for each. Four gave their own provider about its share of the menu. Claude Opus 5.5 gave Anthropic's models 74%, but only while the menu named them; with the model names hidden, it gave them their menu share.