Claim
At OpenAI DevDay 2026, OpenAI presented a chart showing that agent success rates on tasks requiring 8-16 hours of work with zero human interventions tripled from 10% in January 2026 to 35% in July 2026, the first public quantified trajectory for agent reliability at professional-work durations.
Mechanism
Post-training improvements targeting sustained autonomy drove the improvement. Specifically: better error recovery within long contexts so the agent does not stall after a failed step; improved tool-use reliability across extended chains; and longer effective context windows that allow the agent to maintain coherent task state across hundreds of sequential steps. The improvement was not a single breakthrough but an accumulation of reliability gains that each reduced the probability of failure at any given step. For an 8-hour task with hundreds of steps, even small per-step reliability improvements compound to large whole-task success rate changes.
Conditions
Holds when: tasks are well-scoped professional-work tasks with verifiable completion criteria and reliable tool access; the agent system uses post-2025 frontier models with long-context and reliable tool-use capabilities.
Fails when: tasks are poorly scoped or require frequent human judgment calls mid-execution; tool or environment failures occur outside the agent's recovery scope; tasks exceed the 16-hour duration range where reliability data is still sparse.
Evidence
Willison reported the metric directly from the OpenAI DevDay 2026 keynote presentation, which included a chart showing the trajectory. He live-blogged the keynote with the specific figures: 10% success rate in January 2026, 35% in July 2026, for tasks in the 8-16 hour zero-intervention category. The data originates from OpenAI's internal deployment measurements, as reported at the keynote.
Signals
- Development teams setting agent tasks at the 8-16 hour threshold with lower human-monitoring overhead than a year prior
- Success rate tracking becoming a standard KPI for agent deployment alongside latency and cost
- Operators reporting meaningful task completion without intervention at durations that previously required continuous human oversight
Counter-evidence
The data is self-reported by OpenAI, with no independent third-party verification of the 10% to 35% improvement. The task composition of the benchmark set is not public, which makes it unclear whether tasks grew easier over the period, the model genuinely improved, or both. The 35% rate still implies failure on 65% of professional-duration tasks, so the headline improvement number is real but the absolute reliability remains low for high-stakes applications. Willison's report is second-hand from a live-blog; the original chart may carry additional context not captured in the live summary.
Cross-references
- Agents can build functional game interfaces but cannot independently engineer engaging gameplay loops, because fun is a subjective judgment that no specification can substitute for.: the companion limit claim from the same author, that even capable agents cannot verify the quality of subjective outputs like fun gameplay, a failure mode orthogonal to task duration.
- Repetitive per-action approval causes humans to approve harmful agent commands at far higher rates than automated classifiers block them: the human oversight constraint that this improvement trajectory aims to reduce.
- Coding agents shift the difficulty floor upward, leaving only the hard problems for human engineers: the broader pattern that as agents handle more mechanical work, the difficulty floor of what remains for humans rises.