Did you know that we have LinkedIn, Instagram and X accounts
that you can follow?
Hi {{first_name|everyone}},
OpenAI says its researchers now average almost six AI agent workdays in a single day, and the humans are still the ones making the big calls. Here’s the rest of what we covered:
A take on why verifiable code might matter more than readable code once agents are doing the reading
3 new tools worth trying
How readers answered last week’s poll
A music machine where the beat controls the visuals
🤿 DEEP DIVE
OpenAI Is Using AI to Build AI
Coding agents are taking on more research work, but humans still make the big calls.
At OpenAI, a researcher can have several AI agents working at the same time. Some write code, others monitor experiments, and some analyze results. And apparently, it's becoming a pretty normal way to work.
OpenAI released an update showing how much its own researchers rely on coding agents to develop newer AI models. The numbers are quite something.
From March to October, the success rate for tasks that would normally take a human two to four hours increased from 23% to 46%. For longer tasks estimated at 4 to 8 hours, it climbed from 13% to 35%. These are estimates of how often agents completed delegated tasks without human intervention.
So nearly half of those shorter research tasks can be completed without someone stepping in.
Researchers are running several AI workdays at once
OpenAI measures something called an agent workday, which simply means eight hours of an AI agent working.
Earlier in the year, researchers averaged 0.5 agent workdays per day. That figure eventually reached 5.7 agent workdays per researcher per day, an elevenfold increase.
Some teams working on AI training averaged roughly eight agent workdays in a single calendar day. That's possible because multiple agents can work at the same time.
The amount of AI being used is striking, too. At standard API list prices, the median researcher's daily agent usage reached $2,400 in October. For the heaviest users at the 90th percentile, it was around $23,000 per day. Those figures represent the estimated API value of usage, not necessarily actual money spent.
But AI still struggles with some of the work
The success rates vary quite a bit depending on the task.
Agents successfully monitored training runs for problems 79% of the time without human intervention. But when asked to write code for the systems that run AI models, that figure dropped to just 25%.
Researchers also described agents going down unnecessary rabbit holes, producing poor-quality code, or making incorrect assumptions when analyzing results.
One researcher explained that instead of asking an agent to complete an entire batch of work, they ask it to make one example first, review it, and then let it continue.
That sounds much closer to managing a junior colleague than simply pressing a button and getting finished work.
Who actually comes up with the ideas?
OpenAI also asked researchers where the important ideas behind GPT-6 Astra and GPT-6.2 Astra came from.
For the ideas examined in posttraining and safety monitoring, researchers credited humans with coming up with them. AI helped put those ideas into practice.
There were some exceptions in GPT-6.2 Astra's pretraining work. Researchers described developing ideas together with Codex, including cases where it suggested changes or identified useful insights.
So AI is starting to contribute to the thinking itself, although the report still places people at the center of the research process.
And there's another detail worth mentioning. OpenAI reported a 77% increase in documented Safety and Alignment workweeks from January to September, reflecting both new hires and researchers shifting toward safety projects.
OpenAI calls the broader process recursive self-improvement, where AI helps develop future AI systems. The company is tracking whether this is actually accelerating model progress, while acknowledging that more agent activity doesn't automatically mean better models.
For the moment, the picture is pretty clear. AI agents are doing a growing share of the work involved in building AI. But deciding what research is worth pursuing, checking whether the results make sense, and keeping the process under control still require substantial human involvement.
The quick take
Success on two-to-four-hour research tasks rose from 23% to 46% between March and October.
Researchers averaged 5.7 agent workdays per calendar day, up from 0.5.
Agents performed well at monitoring experiments but struggled with more complex coding tasks.
Researchers still originated most of the important ideas examined, although Codex contributed to some GPT-6.2 Astra pretraining ideas.
Documented Safety and Alignment workweeks increased 77% from January to September.
🤝 FROM OUR FRIENDS
A quick share we thought you might find useful
Usage-Based Pricing Is Here. Is Your Finance Team Ready?
More B2B companies are moving to usage-based and hybrid pricing — and finance teams are feeling it. Revenue recognition gets messier, forecasting gets harder, and the manual work compounds fast.
Tabs and PwC teamed up to break it down. In this on-demand session, Rebecca Schwartz and Amit Dhir share how leading finance teams are handling the operational reality of dynamic pricing models — and where AI fits in.
Watch the recording for concrete examples, a practical rev rec framework, and a clear-eyed look at what it takes to scale without the overhead.
If your team is navigating this shift, this session is worth your time.
🗳️ FRIDAY POLL
Where the community stands on AI, tech, and productivity
How good do you think you are at spotting AI images?
Last week’s results:
What do you use AI for most often?
50% - Research and learning
33% - Writing and creating
17% - Others
0% - Organizing work and everyday tasks
🔥 WHAT’S THE BUZZ
What people are building, sharing, and talking about
Developer and educator Matt Pocock argued that the rise of coding agents changes what “readable code” should mean. Humans traditionally favor code that is easy to understand because humans need to maintain it. Agents can inspect far more code, run tests and probe behavior directly, so Pocock argues that verifiability may matter more than readability in an agent-heavy workflow.
The broader question goes beyond programming. Many workplace processes are designed around what makes information easiest for humans to consume. If AI becomes the first reader of documents, data or workflows, we may start organizing some information differently while still keeping enough clarity for humans to step in when necessary.
⚒ TOOL SNAPSHOTS
Futuristic tools within AI, no-code, and productivity
🧩 Netlify Server
Add a backend that stays alive while people use your app.
Why it’s useful
It keeps WebSockets open and state in memory so realtime features can run without dropping between requests. Handy for multiplayer games, AI agents that need time to finish, and per preview environments with their own server.🗺️ Waldo
Turn geospatial questions into answers, maps, and data pipelines.
Why it’s useful
Use it when you need to analyze satellite imagery and geospatial data and get results back within hours. If it needs to keep running, it can schedule the work as a pipeline on your existing data platform.📈 Adside
Manage paid ads across multiple platforms with a team of AI agents.
Why it’s useful
Handy for marketers and small teams that need help creating campaigns, adjusting targeting, and monitoring ad performance across Google, Meta, LinkedIn, Reddit, and X without managing everything separately.
👀 ON OUR RADAR
One more thing we thought you'd like
Turn your expertise into working software
Describe what you want to build
Get real software, not a prototype
Lovable handles payments, logins, databases
Change any detail yourself, any time
Put it in front of customers this week
Discount of $5 applies to first month of any paid plan. You need to sign up via this webpage in order to claim the discount. The discount applies to new and returning users without an active paid plan. Terms apply.
🤖 AI MADE THIS
Interesting and inspiring creations made with AI
A Music Machine Where the Beat Controls the Pictures
Dr. Apolotary used Claude Fable 5.2 and Opus 5.5 to help build Stitches, an experimental music machine where the rhythm also controls the visuals. It combines sounds, musical patterns, and changing scenes into one interactive instrument. It's an interesting example of AI helping someone create a new way to make audiovisual art.
AI used
Claude Fable 5.2
Claude Opus 5.5
Muse Wonder Full
ℹ️ ABOUT US
The Intelligent Worker helps you to be more productive at work with AI, automation, no-code, and other technologies.
We like real, practical, and tangible use-cases and hate hand-wavy, theoretical, and abstract concepts that don’t drive real-world outcomes.
Our mission is to empower individuals, boost their productivity, and future-proof their careers.
We read all your comments - please provide your feedback!
Did you like today's email?
What more do you want to see in this newsletter?











