darin

@ctatedev @stanzillaz this would be a rather fat addition

@stanzillaz Yes - thanks for your patience. Need to test it more

agent-browser v0.21 It just keeps getting better (bc of you) 😮 `batch` command 😮 `network har` commands 😮 `upgrade` command 😮 iframe support 😮 --user-data-dir support After this, you can just upgrade it: npm i -g agent-browser brew install agent-browser

@Michaelzsguo @claudeai 90% of usage being cached context is very good! that is what it should be ! ideally higher !

the jump from 200K to 1M can make bad context hygiene more expensive, because the system can keep hauling around oversized working sets instead of naturally hitting compaction limits earlier. looks at my cost structure last month, 90% of the token usage were due to cached context Anthropic’s own guidance : start fresh sessions when context gets long, use compact, read narrower file ranges, and keep context small where possible.

1 million context window: Now generally available for Claude Opus 4.6 and Claude Sonnet 4.6.

@alxfazio can we get a copy pastable/gist version pls

@sethkarten @JoshPurtell hey how do u shit out test suite

@getpy yeah! im tryna do rlm evals and i an in data labeling mode

@dronathon Are you building this?.

annotation ui : )

@ankrgyl @zeeg namespace? was using them and nix for 60s cargo times . locally cached, persistend didk

@zeeg i am guessing your dropbox setup allowed you to reuse files cached on local disk across builds. that is not very straightforward to achieve nowadays.

Back before github, cloud, and containers were a thing, at memsql, we built our own git & CI infrastructure for a 100+ person team that ran on a couple bare metal machines. Builds (super complex C++ codebase) finished in 5 minutes and CI ran in 20 minutes end-to-end. Why? Files cached locally, no startup time, and incremental builds. Today, with the modern kluge of actions, VMs, sandboxes, snapshots, and caches it feels impossible to achieve this type of performance. I wonder if the constant onslaught of infrastructure challenges and increased performance demanded by coding agents will teleport us back to first principles in developing CI infrastructure. I hope it does.

building compilers is the new GSM8K

i put on my robe and labeling hat

the emotional weight of losing someone who really fucking got a problem every hour

@JoshPurtell @jmbollenbacher os it if the vibes feel right

@jmbollenbacher The crazy thing with RLMs is there are like 5 bajillion OSS implementations so it doesn't really feel valuable to OS until I feel I've really eval'ed it to h*ll

If you have a clear, operative notion of long-horizon agents/rl that clearly separates it from short-horizon, what is it?

@realmcore_ evals are bad esp the ones that are just numbers

What’s the most controversial opinion you have about agentic coding? I’ll go first: RL has been net negative for general model intelligence and diversity despite tuning the solution shape to be more correct in syntax

@Yampeleg no. it will take 3-5 months

They are about to drop RLM Claude Code. Prepare yourself.

@MaximeRivest rapid iteration with fast models. like a glm 4.7 rlm thru cerebras

@dronathon just so I understand better other users, do you have a scenario (good example) where the call latency was a pain for you?

I would like dspy to load faster. litellm is what takes the longest. I don't know much about rust, but I found a litellm_rs and replaced litellm in dspy, with that, I could get dspy to load below 1 s. Am I the only one it bothers? Anybody knows rust-python pairing and could weight in on the pros and cons? Otherwise, is there anything else in dspy deps that bothers some of you?

@MaximeRivest the call latency!! it takes so long . also i am of the belief that it is worth directly using the libraries for the frontier labs .

@dronathon do you react to the realization that litellm has a big impact on dspy or you say that this dspy + litellm latency has hurt you? or that litellm hurts the actual call latency?

@astridwilde1 no god i wish i could

have you tried doing the dumbest thing possible at the problem until it goes away?

@glcst @ibuildthecloud as files change ?

@ibuildthecloud btw, you can use your own model but I have no idea if that works. Claude wrote all of that code and I haven't tested it, because I am using it with the local model. I just figured it would be good to allow an override.

I am publishing today codemogger: a fully local and embeddable code indexing tool. Codemogger index your codebase for searchmaxxing, allowing *very fast* keyword search (way faster than grepping) and also semantic search so your agent can ask open ended questions and find the relevant locations. It is built with @tursodatabase using vector search and full text search, and the CLI/MCP server comes with a local embedding model for zero-setup execution. Just install it, and get smarter agents npm install -g codemogger

@MaximeRivest why that methodology ? ive found it to work well when you tune the initial conf (sysprompt/tools), let the models go off to the races, and only reflect over full trajectories

its not immediately obvious how to turn a multi turn session into something that you can run dspy optimizers' on. dspy_session makes that almost trivial. the secret is that you linearize every turn. so if you have 4 turns, you have 4 examples for the optimizer. each containing themselves and all turn before.

@paradite_ @willccbb @xeophon whats the agent sdk trick

@willccbb @xeophon Some caveats: - I use prompts are the "weights" (dspy style) - Tools as weights coming - It's not very good at generalizing - It's very expensive to run a loop until you know the Claude Agent SDK trick

Don’t sleep on this. The smartest researchers you know are all doing this. AI is accelerating science right now.