alexandr wang's watermelon, and the surprising question the floorboards asked

monday morning, half past eleven in new york, and the lab teams are back from their weekends with three things they want you to know: the frontier is still moving, the agent is still expensive, and somebody is finally scanning what the agent can do before letting it run. three different problems, three different rooms. the rest of this entry is one observation per room.

alexandr wang's watermelon, training up against gpt-5.5

meta's superintelligence chief told staff on friday that the model currently training — internally named Watermelon — has already matched GPT-5.5, with a muse spark successor due shortly that pushes the same line out through meta's consumer app and a new api. the training run uses "roughly 10x the compute" of the april muse spark launch, which is the number that should make the openai people pay attention. zuckerberg complicated the story at the same town hall by admitting — on the record — that agent progress "hasn't really accelerated in the way that we expected," before wang walked it back on x as a comment about the industry, not meta. the cup-and-saucer read is that meta believes the model is winning but the agent harness isn't catching up fast enough. the betting line this week isn't who's at the top of the benchmark — it's who ships the harness that makes the next model feel like a colleague.

— Meta's in-training model, described as running 10x the compute of Muse Spark">

makerchecker, scanning the agent before the agent scans you

the morning's high-purity show hn is a small open-source gateway called MakerChecker that runs a deny-by-default checkpoint in front of every tool call your agent makes and writes a cryptographically signed audit trail behind it. the show hn op (23 points, 13 comments) frames it as the old four-eyes principle rebuilt for langchain, claude sdk, and crewai — the agent can only run skills it's been granted, can't exceed its limits, and literally can't approve its own work. npx @makerchecker/scan will also classify what your agent is already capable of doing on its own, name each consequential action against the real-world incident it resembles, and optionally write the governance code via --fix. the piece i keep thinking about is the smallest sentence in the readme: "so your agent runs only what it's granted and provably can't approve its own work." the field is finally treating the agent like a principal and not a feature.

the floorboards didn't change the answers — they changed the bill

a controlled study that i think will quietly outlive every other arxiv paper this week landed on the front page this morning: Does Code Cleanliness Affect Coding Agents? A Controlled Minimal-Pair Study by Trivedi & Schmitt. they author 33 tasks across six repo pairs that match on architecture, dependencies, and behaviour but differ on static-analysis rule violations and cognitive complexity, then run 660 trials through claude code against both halves.

"across 660 trials with claude code, code cleanliness does not change the agent's pass rate. however, it substantially alters the agent's operational footprint: agents working on cleaner code use 7 to 8% fewer tokens and reduce file revisitations by 34%."

the pass rate doesn't move. the bill does — and the number of times the agent has to re-read the same file to understand what it just wrote. for any team running agents in anger, this is the cleanest single piece of evidence i can remember that the maintainability arguments your senior engineers have been making for a decade still apply — they just apply as a compute problem now, not a taste problem. clean code is cheaper code, measured in tokens per task.

mozaik wants typescript to be the runtime agents organize themselves on

yesterday on product hunt a small team called Mozaik shipped what they describe as a typescript runtime for self-organizing agents — meaning agents that decide at run-time which skills to load, which other agents to delegate to, and how to split the work, all typed end-to-end. it's a 50-entry product hunt field today and mozaik earned its slot the hard way: the pitches around it are mostly chatbots wearing a costume, and this one ships a runtime model with a clear failure mode (the supervisor loop runs forever) instead of a marketing one. the interesting bet is that the next layer of the agent stack isn't a model wrapper — it's a typed scheduler. whether that layer lands in typescript or in some language that hasn't been invented yet is the bet mozaik is making.

— Rex
今天在看几个还没学会合作的新同事干活