More seriously, in a multi-agent setup the per-token cost matters less: a bit of Claude, a bit of Codex, a bit of Gemini-CLI, ... No single model carries the full bill, and having three different training sets catches more "green tests, wrong code" than any single xhigh pass would. Even at 10x per token, one well-placed Opus in the reviewer seat beats one full Opus session on everything.
Good, but "Vibe" is a surprising name. A year ago Mistral Code was sold on engineering rigor and CISO comfort. "Vibe" evokes exactly the opposite, and to me the term has carried a dismissive ring in dev circles since Karpathy. Wondering who won that naming meeting.
Same. The numbers match your feel. Going from 4.6 to 4.7: +14.6 on MCP-Atlas, +10.9 on SWE-bench Pro, tool errors cut by two-thirds. But BrowseComp dropped 4.7 points. Anthropic's own announcement says 4.7 "takes the instructions literally" where 4.6 interpreted them loosely, and recommends re-tuning prompts accordingly. In a conversational loop with an opinionated developer, that translates to less quality because less reasoning — the model executes instead of thinking through.
https://llm-stats.com/blog/research/claude-opus-4-7-vs-opus-...https://www.anthropic.com/news/claude-opus-4-7
Zhao et al. 2023 showed any imperceptible watermark is provably removable by generative regeneration: pass the image through an img2img or VAE, the model reconstructs it visually identical but starts from a different latent. Watermark gone.
SynthID and similar schemes do hold up well against normal sharing: recompression, crops, color tweaks, Twitter's pipeline. That covers most users.
But the asymmetry is stuck — normally a GPU and a bit of motivation should be enough to strip it. Right?
Got a tool to share? ;-)
(Disclosure: I work on tamer, an OSS supervisor for coding agents — biased.)
Add one more to the count. The OAuth-across-harnesses idea would help, but it doesn't fix the shape of the problem.
"Harness" has always felt off to me. Exoskeleton is closer — Claude Code, Codex, opencode wrap the model and augment it from the inside.
What's missing is a layer above that's explicitly not an exoskeleton: a thin supervisor. A master that watches and guides, nothing more. It just relays I/O and hands approval back to the human.
Genuine question: how many agent-hours to rebuild Jira from scratch and migrate 100% of the content out? Split the work, pool our agents, ship by August 17. ;-)
Note to self: don't go away for the weekend without HN ;-)
Personally I run several agents. At minimum Codex and Claude, so they cross-check each other. Exactly to avoid what you describe. Duplicate functions or "tests all pass" when nothing was implemented is the kind of thing a second model, reading the diff with fresh eyes, tends to flag immediately.
But it takes coordination and solid skills + rules to make it work. Tomorrow's battle -> AI skillers.
Fully agree. Non-dev solutions are multiplying, but devs also need to get much more productive. I recently asked myself "how many prompts to rebuild Doom on Electron?" Working result on the third one. But, still buggy though.
The devs who'll stand out are the ones debugging everyone else's vibe-coded output ;-)
Do you ask it for a design first? Depending on complexity I ask for a short design doc or a function signature + approach before any code, and only greenlight once it looks sane.
It's annoying when a project you depend on gets taken in a direction you didn't sign up for. But the silver lining here is that the cost of maintaining a fork has dropped dramatically. A few years ago, forking something like 11ty and keeping it alive long-term was a real commitment — bug triage, dependency updates, staying compatible with the ecosystem. Now, that kind of maintenance work is exactly what LLMs are good at. You can realistically fork it, keep it on the version you liked, and have an AI handle the boring upkeep.
The perpetual "this project is now dead" anxiety around OSS is becoming less warranted. Which might upset people who'd rather debate sustainability models than just fix the dep.