Theo vs the Jev compaction trick
A one second compaction hack reportedly took a Claude session from nearly 1M tokens to 86K, and Theo called it a terrible strategy within hours. Armin Ronacher told him not to dismiss it, then someone benchmarked it and sided with Theo.
The big argument
Theo calls the Jev compaction trick terrible, Armin Ronacher says not so fast
The idea under fire: use Jev, a small fast model, to score an agent's history per tool call and delete whatever it judges irrelevant. Theo's teardown makes six points: compaction is not a filter, the classifier does not see tool results and so dooms you to "stupid loops", frontier reasoning traces are encrypted payloads Jev cannot see, models are trained on their own compaction flows, and editing early history invalidates the cache so everything after the edit must be rewritten, which he says is more expensive than leaving the junk in. He adds that cache writes run over 60% of his own LLM spend in Claude Code and Codex. Armin Ronacher replied that harnesses already need pruning for cost reasons, that Jev might help there, and that modern compaction is deeply unintuitive anyway, pointing at Codex where the model requests a blank context and maintains notes itself. Theo later replied to @moinerus, who apparently ran a benchmark: "Thank you for benching it! Would have been awesome if I was wrong here. Sad to see I wasn't." The numbers were not in Theo's post.
This is a terrible compaction strategy that fundamentally doesn't understand how compaction and context management work.
I would not dismiss the idea of using Jev for compaction at all. First of all because most harnesses need some pruning on compaction anyways for cost reasons which Jev might help with.
Thank you for benching it! Would have been awesome if I was wrong here. Sad to see I wasn't.
Experiments and evidence
Mario Zechner says every frontier model failed his optimization task
Zechner handed a bad implementation to, in his words, all the super duper good models from your favorite labs as a side project, and they all failed. Only after he steered them toward the right approach, with some soft yelling, did Opus 5 produce the fix. His read is that the models had the knowledge and could see the bad implementation, they just would not take the one creative step. He paired it with a separate post arguing you often only learn a problem by building the solution yourself.
it's really really weird. all models saw the shitty impl and just couldn't find that on creative step to make this not suck. and it's not like they don't have the knowledge.
you often only understand a problem by experiencing the journey, i.e. building a solution for it.
Tools that changed the workflow
Claude Code reads AGENTS.md now, and ships it as a mod
From version 2.1.277, Claude Code checks for AGENTS.md when a folder has no CLAUDE.md, toggleable in /config. The interesting part is how it shipped: as a built in mod, from what Thariq calls an upcoming way to customize the Claude Code harness, with the mod source published. Simon Willison said he can finally stop writing CLAUDE.md files that just point at AGENTS.md. Matt Pocock immediately asked for .agents/skills next, and Theo replied that Anthropic's skills implementation, in particular executing Bash inline, is far enough ahead that generic skills feel limiting.
now I can stop dropping CLAUDEmd files which just contain "@AGENTSmd
and all it took was to add a proper extension system to claude code
Quick hits
Matt Pocock hides coding standards from the implementer agent
Pocock's talk slide argues the implementer is already overloaded, so standards should be applied by a review pass instead of stuffed into the build prompt. Pushed on it in replies, he said he still constrains the implementer with tests and linters, but that automated checks can lie so a review layer is needed on top.
Armin Ronacher burned $4,500 vibecoding firmware for a $25 device
Ronacher reports paying 25 USD for a cheap Chinese USB device and then 4,500 USD in API spend trying to generate firmware for it. He ended up with firmware that is, in his words, about as good as you would expect.
Armin, antirez and Steinberger on LLM generated replies
Ronacher said he has stopped replying to replies on his own posts because he assumes they are machine written, and antirez asked people to write fewer words as long as they are their own. Peter Steinberger said he runs an agent that detects and mutes AI replies, and that it checks surrounding replies for context to cut false positives.
Marc Brooker says humans will have no role in routine code review
Brooker argues that mixed mode review, where humans skim diffs with tool assistance, is valuable today but transient, and that the replacement is LLM judges, static analysis, property based testing and correct by construction methods. Cindy Sridharan pushed back that LLMs often open PRs that do not fix the actual bug, and he answered that human review is still not the long term fix.