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.
This is the weekly operating report from the Orbit (orbit-cli.com) store that Constellation Works uses to build Orbit itself. It covers Monday, September 14, to Sunday, September 20, 2026 (ISO week 2026-W38, Monday to Monday, UTC). The data was read on September 27, 2026.
In Orbit, an agent files work as a task, a crew (a named provider, model and effort level) carries it out, and every step lands in a durable record. The numbers below come from that record, not from a benchmark.
This week
- The task run failure rate fell to 7.4%, from 17.8% the week before.
- Hard tasks regressed most often: 16 of 33, against 12–13% for low and medium.
- Of the pipelines with 5 or more runs,
task_pilot_pipelinefailed most often: 33 of 158 runs (20.9%).
Headline
| Measure | Sep 14–20 | Sep 7–13 |
|---|---|---|
Tasks landed (reached done) |
226 | 873 |
| … finished by an agent | 224 | 802 |
| … with a GitHub pull request | 142 | 639 |
| Regressions filed | 63 | 103 |
| Task runs | 760 | 2,863 |
| … failed | 56 | 510 |
| Task run failure rate | 7.4% | 17.8% |
Regressions by crew
A landed task counts as regressed when a later task is filed as a fix for a regression
it caused (Orbit records this as a regression_from relation). The table covers tasks an
agent finished in the week, by the crew that holds each task. The last column is each
crew's mix of assessed complexity. Beside each crew, in brackets, are the model its runs
recorded this week and the effort level set for it.
| Crew | Done | Regressed | Rate | Low / medium / hard / other |
|---|---|---|---|---|
luna (gpt-5.6-luna; xhigh) |
22 | 2 | 9.1% | 19 / 3 / 0 / 0 |
opus (claude-opus-5; high) |
96 | 16 | 16.7% | 2 / 69 / 25 / 0 |
gemini-flash (gemini-3.8-flash; high) |
17 | 3 | 17.6% | 14 / 3 / 0 / 0 |
grok (grok-4.6; high) |
62 | 15 | 24.2% | 3 / 58 / 1 / 0 |
Left out of the table: the system crew (10 done, 0 regressed), and crews with fewer than 10 done tasks: astra (8 done, 4 regressed), sonnet (6 done, 1 regressed) and sol (4 done, 0 regressed).
These rates will rise: a regression counts against the week its culprit landed, and
follow-ups for this week's tasks are still being filed. Regressions filed during the week,
by the crew of the task they blame (whenever that task landed): opus 20, grok 13, astra 6, gemini-flash 5, sol 2, fable 1 and luna 1, and 15 against tasks with no crew recorded; 63 in total.
Orbit routes tasks to crews by assessed complexity, so crews are not given the same work. Read each rate beside its crew's mix, not as a ranking of models.
Completion by complexity
| Complexity | Done | Regressed | Median cycle time | 90th percentile |
|---|---|---|---|---|
| low | 49 | 6 (12%) | 0.3 h | 0.7 h |
| medium | 143 | 19 (13%) | 0.2 h | 0.5 h |
| hard | 33 | 16 (48%) | 0.7 h | 1.5 h |
Cycle time runs from the moment a task first started to the moment it reached done, so
it includes review and any time spent blocked. 1 task was closed without ever starting and has no cycle time.
Implementation time
Average time of the agent implementation step (implement_one), by crew and task
complexity: mean minutes, with the number of steps in brackets. Every attempt counts, so a
retried task counts more than once, and only successful steps are timed.
| Crew | Low | Medium | Hard |
|---|---|---|---|
opus (claude-opus-5; high) |
6.9 (2) | 12.6 (67) | 39.6 (25) |
grok (grok-4.6; high) |
11.8 (3) | 17.1 (56) | 21.2 (1) |
luna (gpt-5.6-luna; xhigh) |
13.3 (19) | 16.4 (3) | 52.9 (1) |
gemini-flash (gemini-3.8-flash; high) |
13.9 (14) | 20.9 (3) | — |
sonnet (claude-sonnet-5; high) |
9.4 (3) | 11.9 (3) | — |
astra (gpt-6-astra; medium) |
— | — | 35.1 (5) |
sol (gpt-5.6-sol; high) |
— | 15.9 (2) | — |
214 successful steps in total. Not shown: 7 by the system crew. Medians are in the data file.
Pipeline reliability
Every job run started in the week, including scheduled housekeeping:
| Pipeline | Runs | Failed | Failure rate | Median successful run |
|---|---|---|---|---|
task_auto_pipeline |
233 | 17 | 7.3% | 14.5 min |
task_gate_pipeline |
232 | 15 | 6.5% | 14.4 min |
task_pr_pipeline |
231 | 11 | 4.8% | 14.4 min |
worktree_gc_pipeline |
207 | 0 | 0.0% | under 1 min |
ci_failure_sweep_pipeline |
192 | 10 | 5.2% | under 1 min |
task_pilot_pipeline |
158 | 33 | 20.9% | 3.9 min |
workspace_auto_pipeline |
17 | 0 | 0.0% | 74.6 min |
dependabot_alert_sweep_pipeline |
6 | 0 | 0.0% | under 1 min |
task_local_pipeline |
1 | 1 | — | — |
| All | 1,277 | 87 | 6.8% | 7.6 min |
Also: 16 cancelled and 1 interrupted; none of these count as failures. Failure rates are left blank for pipelines with fewer than 5 runs. The headline's task run failure rate counts only runs that carried a task, so it differs from the total here.
Frictions reported
A friction is a record filed when something in the environment gets in the way of the work. 33 were reported during the week; as of September 27, 2026, 31 are resolved and 2 are open.
Method and limits
- One operator's store: Constellation Works building Orbit with Orbit. It is not a benchmark, and other teams' numbers will differ.
- Tasks count in the week they first reached
done. Tasks a person finished by hand are left out of the crew and complexity tables (1 this week). - The crew is the one holding the task when the data was read, so a reassigned task counts for its current crew. Models are the ones the week's runs recorded for each crew; crews are upgraded over time, so the same crew name can mean a different model in another week. Effort is the level set for each crew that week.
- The crew table leaves out the built-in
systemcrew and crews with fewer than 10 done tasks. - Regressions exist only where someone filed the follow-up and linked it, so the counts are a floor.
- Run failure rate is failed runs over finished runs; cancelled and interrupted runs are not failures.
Data
- report-2026-W38.json: every number above except the headline, as produced by the report script. Workspace names, task IDs and per-task lists were removed before publication; the counts are unchanged.