Human interventions on 4–8 hour agent tasks (self-reported)
Share of successful four-to-eight-hour agent tasks inside a frontier lab's research organisation that needed at least one human intervention. Today only OpenAI reports it, about itself, and only as 'over half'.
Reason through thisThe reading
Over half of successful long tasks still needed a person in the loop in the first half of 2026; agent effort is up, autonomy on long tasks is not yet.
1 reading drawn from openai_blog.openai.rsi_intervention_rate_4_8h.pt; 1 observation in the evidence.
Source: OpenAI research posts. OpenAI.
The numbers
| as of | reading | value | record |
|---|---|---|---|
| 2026-07-31 | openai (disputed) | 50.0% | reported |
Why it matters
The reliability side of the return arrow. Agent-workdays can rise while humans still steer every long task; the loop only substitutes for people when this rate falls.
What cuts against it
Self-reported and rounded; counts successful tasks only, so failed runs that needed no intervention are excluded; 'intervention' is defined by the lab.
How it is measured
- series
openai_blog.openai.rsi_intervention_rate_4_8h.pt - source OpenAI research posts · default tier 7 · OpenAI; short quotation
- proxy deployment, behaviour · cadence quarterly · valve return arrow
Humans steering most long tasks (half or more need intervention) is the normal-technology picture; a quarter or fewer is the regime where long agent runs are autonomous. Between is emerging. The reported value is a floor ('over half'), stored as 0.5.
Applied to openai_blog.openai.rsi_intervention_rate_4_8h.pt.
Latest readings
- 2026-07-31 openai · 50.0%⚑disputed: OpenAI measuring its own research organisation and publishing the result itself; no independent check (tier 7).as of 2026-07-31
Why this status, and its history
Same OpenAI post: over half of successful 4-8 hour tasks in the last six months involved at least one human intervention, stored as the floor 0.5. That sits at the edge of the normal band (half or more), and tier 7 caps the status at emerging regardless. Humans still steer most long agent runs. Initial seed.
- 2026-09-10unmeasured to emergingconf — → 30 · the evaluator (automatic)
Same OpenAI post: over half of successful 4-8 hour tasks in the last six months involved at least one human intervention, stored as the floor 0.5. That sits at the edge of the normal band (half or more), and tier 7 caps the status at emerging regardless. Humans still steer most long agent runs. Initial seed.
source1
Why this confidence
Rests on 1 observation from OpenAI research posts; best evidence tier 7, grade D; current.
Confidence is independent of status: 90–95 multiple strong independent sources; 70–89 good evidence, some ambiguity; 50–69 mixed or hard to operationalise; 0–49 limited or vague evidence.
Predictions that lean on this
- Too early to tellAI speeds up its own invention · This site
- Too early to tell
- Too early to tell
- Too early to tell
- Too early to tell
More predictions that lean on this
- Too early to tell
- Too early to tell