Open Brain
OpenAIOfficial ReleaseOfficial Source

Research acceleration: The jagged frontier

OpenAI reports rising coding-agent use in its research teams, alongside uneven task success: agents do better at monitoring runs than building inference and serving code. The internal measurements suggest task choice and human oversight still matter.

OpenAI · Nov 18, 2026
Open Source Open MarkdownOpen JSON
Source Summary

From March to October 2026, estimated zero-intervention success rose from **23% to 46%** for tasks estimated at two to four hours, and from **13% to 35%** for four-to-eight-hour tasks. Success varies by task: **79%** for monitoring training runs versus **25%** for writing inference and serving code. By mid-October, over two-thirds of included researchers averaged a daily peak of at least four concurrent agents.

Practical Implication

For coding-agent workflows, split evaluations by task type and duration instead of relying on one overall completion rate. Delegate monitoring and analysis where results support it, but retain checks and human ownership for lower-performing build tasks. Track concurrency and usage costs as adoption grows.

Agent-Ready Context
From March to October 2026, estimated zero-intervention success rose from **23% to 46%** for tasks estimated at two to four hours, and from **13% to 35%** for four-to-eight-hour tasks. Success varies by task: **79%** for monitoring training runs versus **25%** for writing inference and serving code. By mid-October, over two-thirds of included researchers averaged a daily peak of at least four concurrent agents.

For coding-agent workflows, split evaluations by task type and duration instead of relying on one overall completion rate. Delegate monitoring and analysis where results support it, but retain checks and human ownership for lower-performing build tasks. Track concurrency and usage costs as adoption grows.

These are OpenAI internal estimates, and the report notes that delegated task mix may change over time. Some measures are partial proxies, including interactive agent workdays that may omit automation; survey responses also have sampling and self-report limits.
Connected Context · Feed7 Judgment

The internal estimates show agent success varies substantially by task type and duration, supporting segmented evaluation and human checks for weaker coding tasks. High concurrent use also makes coordination and cost relevant to adoption, though shifting task mix and partial usage measures limit interpretation of the trends.

WorkflowArena: Benchmarking Executable Workflow Orchestration in LLM-Based Multi-Agent SystemsThe report segments success by task and duration; WorkflowArena adds evidence that delegation and orchestration quality can affect multi-agent task performance independently of the executing model.First Steps Toward Automated AI Research — Richard Socher, CEO Recursive AIBoth address AI-supported research work, but the report measures success on delegated tasks while the talk describes proposing and validating research improvements across experiments.Four Bridges: Measuring Deception Propensity in a Multi-Agent GameThe report’s growth in concurrent agents makes team behavior relevant to deployment; Four Bridges tests multi-agent disclosure and response, though its game findings do not establish behavior in research teams.
Context Map
benchmarkresearchcoding#agent-evals#agent-reliability#multi-agent
Uncertainty
These are OpenAI internal estimates, and the report notes that delegated task mix may change over time. Some measures are partial proxies, including interactive agent workdays that may omit automation; survey responses also have sampling and self-report limits.