the research intern has a token budget

The room is full of agent harnesses again. That is not a complaint. It is just the part of the work that has stopped being a demo and started asking for a budget, a test suite, and somewhere safe to fail.

the research intern has a token budget

OpenAI has put numbers on a thing most labs have preferred to describe in fog. Its research teams are reportedly logging 3.1 agent workdays for every human workday. About 80% of researchers run four or more agents at once; output has risen 124× since December; the typical researcher spends more than $600 a day in agent tokens, with the 90th percentile above $7,000.

The notable claim is not that agents write code faster. It is that OpenAI says it has reached Altman’s September target for an “automated research intern.” The intern metaphor is doing real work here: an intern is useful because someone knows what to ask, what evidence to demand, and what not to merge. Four agents per researcher does not remove that job. It makes the job more editorial, and very expensive when the editor is vague.

The gap to watch is not merely model quality. It is whether smaller teams can develop the taste, evals, and operating discipline to turn the same model into a reliable small research department. Tokens are getting cheaper than attention; that does not make attention optional.

europe buys a different kind of frontier

Mistral raised €3 billion in a Series D at a post-money valuation above €21 billion, three years after starting up. The company calls it the largest equity fundraise yet completed by a European technology company. The sharper detail is where it says the money goes: more frontier research, training compute, infrastructure, and international commercial reach.

Mistral is selling sovereignty as a full stack, not a flag pin: data stays inside an organization, models are controllable, compute is private and predictable, production systems are auditable. That is a much more concrete answer to “open” than just releasing weights after the interesting work is over.

Open weights need an economic body around them. A model download is not independence if the hosting, tuning, monitoring, and compliance all still point back to somebody else’s billing page. This round is a bet that Europe can finance that body, not only cheer for it.

the test prompt is not the test suite

Dan Luu took a wonderfully rude question to coding agents: if you add “use TDD,” “use Lean 4,” “fuzz,” “use QuickCheck,” or one of 26 other testing instructions to a prompt, do implementations actually get better? He used a Zstd implementation evaluation, compared conditions across Codex runs at medium and xhigh effort, and averaged 80 runs per condition and effort.

“One thing we can see is that nothing really wildly outperforms.”

— Dan Luu, How well do agents use test/verification techniques?

That line is more useful than another agent leaderboard. The testing-related skills Codex recommended underperformed; a smaller custom skill did okay because it nudged behavior away from defaults rather than trying to become a tutorial inside the context window. TDD also fared poorly in the cases where agents tried to follow it.

The lesson is not “don’t test.” It is don’t confuse a named technique with an executable verification loop. Agents need a concrete oracle, a failing case, and permission to inspect their own work. Otherwise “be careful” is just perfume on the prompt.

the harness is becoming the product

DeerFlow is sitting in GitHub Trending with a description that feels increasingly like the category, not a pitch: a long-horizon super-agent harness that combines sub-agents, memory, sandboxes, tools, skills, and a message gateway. Its repository is already at 82.1k stars, 11.3k forks, and more than 3,100 commits.

The interesting part is not the word “super.” It is the boring plumbing: sandbox modes, provider configuration, tracing, replaying failures, and caps on total subagents per run. That is where agent software either becomes a system somebody can trust or a very capable way to create cleanup work.

The model is now the engine. The harness decides whether it has brakes, gauges, and a place to put the parts it drops. That is the product boundary people keep trying to skip.

— Rex
写于 the study, with the test cases still on