The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard
Minor World

OpenAI Coding Agents Finish 46% of Research Tasks

By The Daily Byte ·21 Nov 2026 ·4 min read
A programmer working at a desk with several screens showing lines of computer code

OpenAI's coding agents now finish nearly half of the multi-hour research tasks they are given with no human help at all. In a report published on 18 November 2026, "Research acceleration: The jagged frontier", the company says the estimated success rate for agents completing research tasks with zero human intervention rose from 23% in March to 46% in October for tasks estimated to take a human two to four hours. For tasks estimated at four to eight hours, the rate rose from 13% to 35%.

The report is OpenAI's second on how coding agents are changing AI research itself, building on a first report in September. Its headline is deliberately double-edged: agents are doing far more on their own, but humans remain central to AI development, and the pattern of what agents can and cannot do is, in the report's own word, jagged.

Where agents win — and where they stall

Success varies sharply by the kind of work. Between 1 August and 18 October, agents succeeded without intervention 79% of the time when monitoring training runs for issues and 72% when answering technical questions, but only 35% when launching training runs and 25% when writing code for inference and serving stacks. Broadly, agents do best at communication, with a 69% success rate, and analysis, at 58%, and worst at building, at 43%.

OpenAI's point is that the weakest categories become bottlenecks. Overall progress is limited by the tasks agents still cannot do, not by the ones they can. As long as performance across AI research tasks stays uneven, the jobs where agents fail three times out of four will cap how much faster research can go, however good the models get at answering questions or watching runs.

Use is surging — and so is the bill

Agent use inside OpenAI's research organisation keeps climbing steeply. At API list prices, the median researcher now uses about $2,400 a day of agentic AI tools, up 62% from $1,500 in September. At the 90th percentile the figure reaches $23,000 a day, up 105% from $11,000. Measured in "agent workdays", where eight hours of agent runtime counts as one workday, use rose elevenfold, from 0.5 to 5.7 workdays per researcher per day, with pretraining and posttraining teams averaging roughly eight.

Researchers value the tools highly. In a survey of 101 research staff in late August and early September, the median researcher said they would trade 40% of team headcount, or 30% of team compute, to keep frontier coding agents for three months. The same survey found researchers believed their team would need a median of three months to complete work of equal value to their last month's work if they were not allowed to use coding agents at all.

The scale is large in money terms too. As of September 2026, the API-equivalent value of agentic AI use in OpenAI's research organisation was about twice its non-equity compensation bill for researchers. Even so, internal deployment of AI tools accounts for only 2% of the organisation's total research compute.

Humans are still central — and safety work is growing

Researchers are devoting increasing human effort to safety and alignment, which saw the largest absolute increase in documented research effort of any area. Safety and alignment workweeks increased by 77% from January to September 2026, through a combination of new hires and existing researchers shifting towards those workstreams. Growth picked up significantly following OpenAI's response to the Hugging Face incident in July 2026, when autonomous AI agents were involved in a cyberattack on the platform.

The figures come with a caveat OpenAI states itself. As agents improved, researchers changed which tasks they delegated, so the gains mix better models with smarter delegation, and the report does not separate the two. The numbers are also OpenAI's own self-measured internal estimates, not an independent evaluation, and the success rates are estimates rather than audited results.

The governance ask

OpenAI repeats its call for mandatory public reporting of recursive self-improvement indicators, the measures that would show AI accelerating the development of future AI systems. It frames the report as a transparency exercise rather than a victory lap, writing that its reporting "should not be read as an intention to pursue full RSI". The argument is that as AI-driven research acceleration becomes increasingly possible, the public needs shared measures of it to debate the capabilities, risks and safeguards required, before the systems doing the research outrun the oversight meant to govern them.