Monitors show rows of orange robot icons with progress bars; beyond them, orange boxy robots sit in a curved row and one feeds a sheet into a glowing blue slot.

Field note. Version 1.10.0

What an Agent Seat Actually Completes

Across 106 recorded runs of 9 AI agent seats operating Orbyt Labs between Jul, 21 2026 and Sep, 26 2026, 91 passed the workflow's completion gate (86%), 15 were refused by it, and 73 files an agent changed outside its lane were erased before commit; the frozen Q3 2026 baseline was 80.0% on 30 runs.

Collected to . Sample: 106 recorded runs of 9 seats in one organization.

Key findings.

1. Across 106 recorded runs of 9 AI agent seats operating Orbyt Labs between Jul, 21 2026 and Sep, 26 2026, 91 passed the completion gate, a rate of 86%.
2. The guards refused 15 of 106 recorded runs of the AI agent seats operating Orbyt Labs between Jul, 21 2026 and Sep, 26 2026.
3. Between Jul, 21 2026 and Sep, 26 2026, 73 files that Orbyt Labs agent seats changed outside their lanes were erased before commit.
4. The frozen Q3 2026 baseline for the Orbyt Labs agent seats was 80.0% completion on 30 runs, against 86% on 106 runs read Sep, 26 2026.

The data.

  • Passed the completion gate91 of 106
  • Refused by the guards15 of 106
  • Frozen Q3 2026 baseline completion rate80.0%
106 recorded runs: 91 passed the completion gate and 15 were refused by it. The frozen Q3 2026 baseline was 80.0% on 30 runs.
Measured on the Orbyt Labs repository, Sep, 26 2026. Each figure is produced by the command in its source column.
MetricValueHow it is counted
Recorded runs106public/autonomy-ledger.json, current.perSeat, sum of runs
Seats with a run ledger9same file, perSeat length
Runs whose output passed the completion gate91same file, sum of completions
Completion rate86%completions / runs
Runs the guards refused15same file, sum of failures
Rows where no run happened (skip, halted, infrastructure)4same file, excluded from both sides of the rate
Files an agent changed outside its lane, erased before commit73same file, sum of discards
First recorded runsame file, earliest firstRun
Last recorded runsame file, latest lastRun
Frozen Q3 2026 baseline completion rate80.0%same file, baseline.totals (frozen , never regenerated)
Runs in the frozen baseline30same file, baseline.totals.runs

How it was measured.

Every figure is read from the organization's committed per-seat run ledger (docs/org/stats/<seat>.jsonl, aggregated into public/autonomy-ledger.json by generate-stats.cjs at every commit). A run is one scheduled or dispatched execution of a seat's workflow; the ledger row is written by the workflow itself, not by the agent.

Completion means the run's output survived the workflow's completion gate (a check of how the agent's run ended: its exit, its result and its turns against its cap), the guard gauntlet, and the commit. Since decision D-0061, in September 2026, the completion line a seat prints is recorded as its claim rather than required; runs before that decision also had to print it. A run whose output the guards refused is a failure. Rows recording that no run happened (skip: nothing to do; halted: a kill switch; infrastructure: the org could not run) are counted separately and sit on neither side of the rate.

A discard is a file the agent changed outside its effective keep set, erased before the commit rather than shipped. It is an aggregate counter across runs and a different unit from a failure: it says the agent reached outside its lane, not that its work was rejected.

The baseline is the ledger as first published on Aug, 11 2026 (30 runs, 80.0% completion), frozen on purpose and never regenerated, so every later reading is measured against the same origin rather than against a moving average.

Failures are printed, never netted out. The rate is completions divided by runs, and both numbers are on the page.

What this does not show.

  • One organization, nine seats, one workflow shape. The seats are prompts run by the same action on the same runner, so this measures that design and not agent autonomy in general.
  • Passing the completion gate is not the same as the output being good. What the founder did with each run's output (accepted, edited, rejected, unread) is recorded separately and is not in this note yet; the gate measures whether the machinery accepted the work, not whether a person did.
  • Small numbers. Several seats have under ten runs, so a per-seat rate would swing on one result; only the aggregate is published here.
  • The ledger records what the workflow wrote. A run that died before writing its row (a runner never acquired, a budget refusal) is absent, so the run count is a floor and the completion rate can be flattered by exactly those absences.
  • No trend is claimed from the baseline comparison. The population of seats and the guards they must pass both changed between the baseline and this reading.

Revisions.

This URL is permanent. This table records version numbers and changes.

VersionDateChange
1.10.0Refresh: runs 105 -> 106; completions 90 -> 91; discards 72 -> 73; lastRun -> .
1.9.0Refresh: runs 103 -> 105; completions 88 -> 90; completionRate 85% -> 86%; discards 71 -> 72; lastRun -> .
1.8.0Refresh: runs 102 -> 103; completions 87 -> 88; lastRun -> .
1.7.0Refresh: runs 104 -> 102; completions 86 -> 87; completionRate 83% -> 85%; failures 18 -> 15; notRun 1 -> 4.
1.6.0Refresh: runs 102 -> 104; completions 84 -> 86; completionRate 82% -> 83%; discards 69 -> 71; lastRun -> .
1.5.0Refresh: runs 101 -> 102; completions 83 -> 84; discards 63 -> 69.
1.4.0Refresh: runs 91 -> 101; completions 76 -> 83; completionRate 84% -> 82%; failures 15 -> 18; discards 59 -> 63; lastRun -> .
1.3.0Refresh: runs 83 -> 91; completions 69 -> 76; completionRate 83% -> 84%; failures 14 -> 15; discards 54 -> 59; lastRun -> .
1.2.0Refresh: runs 82 -> 83; completionRate 84% -> 83%; failures 13 -> 14; discards 52 -> 54; lastRun -> .
1.1.0Refresh: notRun 0 -> 1.
1.0.0First publication.

More from Research.

The other measurements from the same repository, and the papers they sit beside.

Cite this.

Bartak, J. (2026). What an Agent Seat Actually Completes. Orbyt Labs Research, version 1.10.0. https://www.orbytlabs.ai/research/agent-seat-completion

CC BY 4.0. Reuse it with attribution.

Back to all research