Anders Lie

The coding agent big AI doesn't want you to know about

Replaced all useEffect with effect-atom in our opentui react frontend. Probably the first time ever that I feel like I've had a sane frontend codebase. Shoutout @tim_smart. Hits extra different paired with effect RPC queries

@kitlangton man vs clanker

part ii

@julien_c Back in the days when o1 was the odd one out

I still remember life before chain-of-thought reasoning models

@kitlangton incredible, will these be joining the opentui codebase?

OPENCODE II — 𝓛𝓘𝓠𝓤𝓘𝓓 𝓒𝓡𝓞𝓦𝓝

guys we have a pretty substantial opensource zig codebase and i'm terrified he's gonna look at it

@simonklee huh hm ok.. one second, let me just, replace our entire agent's markdown rendering pipeline with this rq

The Proportional Terminal

@kitlangton @opencode pretty cool! though makes you wonder in what cases you might just want to intentionally change the prefix rather than append deltas, e.g. if you are on a provider with known low cache ttl, or if returning to a session after a long time

One feature I got to implement in @opencode 2.0 was allowing for HOT CONTEXT¹ without busting the cache. Here's me explaining how it works ↓ ¹ ⁿᵒᵗ ᵃ ᵗᵉʳᵐ

@0xblacklight half a day per ID changed* obviously

wanted to give sol a try. had it create a markdown plan, told it to implement it directly, it refuses and starts running various commands instead. very off-putting vibes so far i must say

@tejalpatwardhan is this like a mad libs?

GPT-5.6 sol post-trained luna!

@victormustar opus has such an ego, obnoxious honestly, it doesn't even have this in training and just assumes superiority by default

Opus just called GLM-5.2 a "weak model" lol I don't think it's up to date with it 😅

Took this to an extreme a while ago with a quickjs interpreter and had the model write exclusively js. Seems nice at first until it starts trying to chain complicated workflows into one turn/program, instead of perceiving results and acting on them dynamically which is arguably like the whole strength of using an agent in the first place. Also why give it some artificially limiting sandbox when it’s a coding agent that could write a real program with actual access to project specific functionality and real libraries tools etc.

The primary benefits of code mode are that it lets the agent: - use syntax it's well-trained on (e.g., TypeScript) - sequentially compose tools w/o round-trips Neither of these entail a separate isolate or process. You can just write a TS interpreter. Am I missing something?

@KentonVarda Yeah the one thing I still write manually when coding is commit descriptions. If you can’t describe the commit even briefly you probably shouldn’t commit it at all

I just declared a moratorium against AI-written change descriptions (e.g. PR and commit messages, also issues/tickets) from my team. AI was writing change descriptions that were worse than useless to me as I tried to review PRs: outlining details of the code that could easily be seen by looking at the code, but omitting the higher-level framing needed to understand broadly what the code is doing. I think people like having AI write these things because the output looks structured and thorough, which makes it feel professional in a way. But this isn't actually valuable. Concise, high-level descriptions are better for everyone. If I need to use my own AI to interpret what your AI wrote then something is wrong. Let AI write code, sure, but for the description, I'd rather see your prompt than your output. We could maybe have extended agents.md with guidelines on writing descriptions, but this seemed a bit pointless since a good, concise change description only takes a few minutes to write -- not a significant time savings to delegate to AI. At least, it doesn't take long if you understand the code -- and if you don't understand the code, then I'm definitely not merging it.

@0xblacklight agree. For serious work you end up being the thinking bottleneck in either case. And similarly glm 5.2 feels highly comparible to gpt 5.5 in these situations

Hot take and I’m sure y’all are all going to tell me it’s because I’m “holding it wrong” (some have already said this in fact) But on anything that’s decently well-specified and not a YOLO one-shot “how good can it get” prompt or a “run it for six days uninterrupted to build a compiler” type nonsense i.e. on anything where you still have to read the code and that structurally resembles how humans still build real software the difference between fable and GPT-5.5-xhigh is not that big Xhigh does tend to overthink a bit but generally both came up with similar solutions and implementations to problems threw at them Makes me wonder about gpt-5.5-pro which afaik nobody has an opinion on because it’s API billing only and ridiculously expensive so nobody tried it

@charles_irl @grok cmon man, what the hell

new benchmark just dropped

@maximelabonne @huggingface Hopefully gets adopted fast so I can finally remove harness level doom loop detectors..

New training technique to reduce doom loops! We applied it to LFM2.5-2.6B (SFT checkpoint) Qwen3.5-4B. By reducing doom loops, it also improves downstream evals. We open-source the training code and training dataset on @huggingface

@ASvanevik 5. Same flag colors

Now that the US is knocked out, I am formally extending an invitation to the American people to support Norway. Why? 1: The Vikings discovered America before Columbus. 2: There are more ethnic Norwegians in the US than in Norway. 3: Next weekend we can pillage the English peasants together. 4:

@lgrammel I think a lot of it comes down to “does this harness properly accommodate model X and not actively work against it”. Once you’re giving the model tools and an environment it’s comfortable in, the remaining delta is workflow specific prompting etc

what is the evidence that pre-built agent harnesses such as claude code or codex are superior (in terms of outcomes)? e.g. compared to using an open source harness like pi or opencode? or compared to creating a custom harness for a specific task/project?

@liquidai this is great, reasoning on small models is usually just looped nonsense. maybe smaller reasoning models will become usable now

Today we release Antidoom, an open-source method that removes a common failure mode in reasoning models: the doom loop. Doom-loop rates before and after, with eval scores up across the board: > Early LFM2.5-2.6B checkpoint: 10.2% → 1.4% > Qwen3.5-4B: 22.9% → 1% (greedy sampling) 🧵

@vikhyatk why is it eating all the hexagons, stop it

look at what big rectangle took from us