16:[[49172936,{"by":"Mzzzzz","descendants":27,"id":49172936,"kids":[49177318,49176640,49174616,49175456,49173396,49173207,49173318,49178325,49178636,49175653,49175561],"score":38,"text":"$17","time":1785868562,"title":"Launch HN: EdotEnv (YC S26) – Quant Trading RL Envs to Teach LLMs Research","type":"story","url":"https://edotenv.com/"}],[49175422,{"by":"exabrial","descendants":6,"id":49175422,"kids":[49180096,49177759,49180365,49180040,49175652,49180296,49178318],"score":7,"text":"Dear Anthropic, can we please have thought traces back?

I can't verify whether or not the LLM is arriving at the conclusion from cheating, or if it's fudging or making stuff up.

Opus 4.6 remains the best model because of this.","time":1785878890,"title":"Ask HN: Dear Anthropic, can we please have thought traces back?","type":"story"}],[49176440,{"by":"michaelmure","descendants":6,"id":49176440,"kids":[49179553,49177250,49176625,49178332,49179894,49178310,49181093,49177607],"score":5,"text":"For CLI tools, humans like visually organized text possibly with colors or TUIs. Machines likes tabs-separated tabular data, or json and similar.

What's the equivalent for LLMs for tabular or structured data? What makes them grasp the data easily and efficiently?","time":1785884248,"title":"Ask HN: What is a good format for a tool to report data to a LLM?","type":"story"}],[49162887,{"by":"unjuno","descendants":2,"id":49162887,"kids":[49162894,49164565,49165889,49166053,49163777,49163831],"score":3,"text":"It seems that generating hypotheses is no longer the main problem.\nThe harder problem may be continuously collecting real-world data, building accurate world models, and using them to guide experiments.\nIs this where automated research is heading?","time":1785801568,"title":"Ask HN: Does AI research need \"world models\" more than bigger LLMs?","type":"story"}],[49160339,{"by":"iagooar","descendants":3,"id":49160339,"kids":[49160466,49160482,49165877],"score":1,"text":"$18","time":1785785447,"title":"Ask HN: I had Codex and GPT 5.6 Sol running for 12 days, 870k+ LOC. Now what?","type":"story"}],[49158267,{"by":"hairymouse","descendants":1,"id":49158267,"kids":[49158594,49165874,49158464,49163780],"score":7,"text":"My question might seem stupid, but ReAct pattern seems to always work and actually gives great results in the other hand we have the plan pattern and to be honest i cannot think of how it's exactly used in production system, im not sure if this is correct but how it's technically implemented is you give the LLM context and you tell it, hey give me your plan to how you will solve this thing, so now it will build the plan but won't it also follow the same plan in the ReAct pattern to solve that exact problem? I want learn from people that used the planning pattern in production, how and why you used it? and how do you enforce the LLM the follow the plan exactly?","time":1785775815,"title":"Ask HN: ReAct vs. Planning Pattern","type":"story"}],[49145307,{"by":"mopatches","descendants":2,"id":49145307,"kids":[49150285,49151753,49153008,49145731],"score":1,"text":"I want something like the agent sidepanel in VS Code/Cursor that runs in the cloud, not on my laptop. I want to choose the model, chat with it on my way into work or from the couch, and get a PR when it's done.

Cursor Cloud Agents works the best, but I think the rest of the Cursor product sucks. Claude Code Web seems abandoned, GitHub Copilot is buggy and regularly cannot see its own chat history, and products like Devin are so focused on token and model indirection that I don't trust the cost-per-task.

Someone has to have good cloud agents. What are you all using?","time":1785682993,"title":"Ask HN: What cloud agents do you use?","type":"story"}],[49139845,{"by":"skeptic_ai","descendants":17,"id":49139845,"kids":[49139888,49141797,49153344,49140472,49142248,49142116,49141677,49140102,49142259,49165980,49149394,49169698,49152998,49140337,49142585,49148713,49140393,49144469,49146483,49141324,49150587,49142720,49140004,49144314,49145099],"score":17,"text":"I’ve asked AI a few times and I still don’t get it.

Why do frameworks like Claude Code or Codex have the concept of “skills” instead of just using well-organized Markdown docs?

Couldn’t I just have an AGENTS.md that points to folders of .md files and tells the agent when to read them? That feels functionally equivalent to me.

What am I missing? Is there a real architectural benefit, or is it mostly a standardization/convenience thing?

I doubt we’d create so much hype about skills if they’d just be a md file that’s in a skills folder. So I feel I’m missing something.","time":1785629342,"title":"Ask HN: I still don't understand why AI agents need \"skills\"","type":"story"}],[49150433,{"by":"leandrobon","descendants":27,"id":49150433,"kids":[49161237,49150757,49150748,49151233,49157004,49153822,49160507,49161415,49159582,49152159,49151155,49151014,49154517,49155285,49177921,49154421,49177750,49152985,49163782,49153103,49153342,49151532],"score":11,"text":"Have we hit a point where most advances come in the form of new integrations, or applications, rather than outright better models? How much more productivity will we be able to squeeze from them?","time":1785722976,"title":"Ask HN: Have LLMs Plateaued?","type":"story"}],[48908950,{"by":"laalshaitaan","descendants":52,"id":48908950,"kids":[48914496,48910221,48914292,48989594,48912431,48966922,48913804,48946519,48913242,48914609,48910168,48912036,48917610,48916278,48914296,48922973,48909305,48923994,48913314,48915226],"score":85,"text":"$19","time":1784045178,"title":"Launch HN: Agnost AI (YC S26) – Extract user feedback from agent conversations","type":"story","url":"https://agnost.ai"}],[48819808,{"by":"segmenta","descendants":99,"id":48819808,"kids":[48831357,48825074,48824783,48822779,48827471,48831257,48824567,48832111,48823596,48831776,48823130,48836157,48829724,48830035,48825473,48832659,48832222,48823705,48830787,48832433,48830797,48833243,48836761,48825960,48823621,48827550,48989570,48968469,48839291,48821053,48833311,48842948,48829368,48825079,48831302,48826371,48829004,48829613,48831842],"score":219,"text":"$1a","time":1783440644,"title":"Show HN: Rowboat – Open-source, local-first alternative to Claude Desktop","type":"story","url":"https://github.com/rowboatlabs/rowboat"}],[48883305,{"by":"julesrms","descendants":119,"id":48883305,"kids":[48911223,48908982,48909080,48907906,48913718,48917213,48908253,48910692,48909030,48919857,48910245,48910313,48910481,48910225,48913695,48907654,48915739,48910753,48908678,48916658,48908108,48908615,48916308,48915314,48910859,48907935,48917509,48917654,48909345,48917499,48910331,48909021,48908629,48910618,48906597,48918022,48910077,48918747,48910657,49074790,48913137,48989561,48915580,48908656,48918778,48919071,48907340,48908803,48910498],"score":280,"text":"$1b","time":1783880895,"title":"Show HN: Juggler – an open-source GUI coding agent, by the creator of JUCE","type":"story","url":"https://github.com/juggler-ai/juggler"}],[48816724,{"by":"jarodrh","descendants":24,"id":48816724,"kids":[48872619,48873987,48841498,48870323,48881397,48839545,48839736,48855432,48968057,48867590,48843538,48837992,48843451,48876994,48829495],"score":67,"text":"$1c","time":1783426854,"title":"Show HN: Frugon – Find which LLM calls a cheaper model could handle (local, MIT)","type":"story","url":"https://github.com/Rodiun/frugon"}],[48882716,{"by":"brryant","descendants":131,"id":48882716,"kids":[48884370,48885657,48883952,48888523,48889887,48889179,48887729,48884572,48885442,48896985,48886638,48888797,48894767,48887574,48894857,48893207,48887918,48887893,48942760,48886086,48884363,48887548,48887397,48885331,48887326,48886808,48885518,48885521,48884146,48884553],"score":258,"time":1783876387,"title":"Migrating a production AI agent to GPT-5.6: 2.2x faster, 27% cheaper","type":"story","url":"https://ploy.ai/blog/migrating-a-production-ai-agent-to-gpt-5-6"}],[48883343,{"by":"therepanic","descendants":322,"id":48883343,"kids":[48884473,48883823,48883835,48883962,48885569,48883936,48884217,48884091,48886652,48884288,48883731,48883743,48883897,48905614,48884834,48883774,48891118,48884842,48884239,48889800,48886046,48883782,48883833,48884072,48884378,48886835,48885076,48896492,48887215,48888808,48884410,48884353,48885697,48885965,48885045,48884368,48884036,48884957,48887632,48902455,48884183,48891180,48883813,48887543,48888172,48885087,48884261,48883779],"score":501,"time":1783881116,"title":"I love LLMs, I hate hype","type":"story","url":"https://geohot.github.io//blog/jekyll/update/2026/07/12/i-love-llms.html"}],[48883275,{"by":"systima","descendants":396,"id":48883275,"kids":[48883796,48883663,48883533,48883969,48883534,48883768,48883571,48885056,48884712,48883545,48885809,48883684,48883730,48883721,48883549,48883991,48883634,48890380,48883538,48884620,48884459,48883737,48886503,48891653,48887451,48888081,48888388,48885008,48896971,48911315,48899340,48884057,48885532,48892552,48883528,48885007,48886287,48886802,48888226,48897535,48887380,48898801,48884727,48885797,48885211,48890775,48888319,48883520,48889809,48885446,48884869,48883815,48893578,48884334,48884612,48884827,48884148,48896864,48886106,48883996,48883894,48883529,48886840,48883986,48888251,48887596,48885169,48894111,48888619,48883702,48884721,48889181,48893531,48885871,48931453,48954298,48889612,48889987,48890358,48890343,48886071,48886561,48936335,48889023,48887530,48884562,48887736,48924765,48886434,48891661,48890235,48953435,48883973,48885086,48883650,48884179],"score":706,"text":"This started based off of a hunch. We usually use OpenCode, but were 'forced' to use Claude Code for a while due to issues with Meridian. In that time, we saw the usage meter rise much, much more quickly than when using OpenCode.

This was the initial anecdotal evidence, but we undertook this small study to collect empirical data:

We added logging between the agentic coding tool (Claude Code and OpenCode) and Anthropic's endpoint, and captured all requests (and the returned usage blocks).

With one caveat (toward the end of the post) we found unambiguously that Claude Code was far more inefficient in terms of its cache strategy and its harness token usage than OpenCode.","time":1783880751,"title":"Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k","type":"story","url":"https://systima.ai/blog/claude-code-vs-opencode-token-overhead"}],[48854585,{"by":"foxtrot8672","descendants":11,"id":48854585,"kids":[48855026,48854641,48856812,48855961,48855527,48861204,48856953,48854868,48855111,48855010,48857995,48906682,48858040,48857300,48856550,48856769],"score":6,"text":"Would you trust an AI to assist you with financial decision making? What would be some requirements to be comfortable with it?","time":1783645922,"title":"Ask HN: Thoughts on AI in Personal Finance?","type":"story"}],[48853669,{"by":"franze","descendants":10,"id":48853669,"kids":[48853993,48853846,48853857,48853961,48853736,48858611,48853959,48857287],"score":6,"time":1783638890,"title":"Ask HN: Did Google give up in the AI Race?","type":"story"}],[48844345,{"by":"cmar00","descendants":26,"id":48844345,"kids":[48857217,48844586,48865487,48914358,48852331,48872733,48844774,48875694,48866830,48861430,48852439,48953430,48931543,48845822,48857274,48866120,48870514,48872759,48862713],"score":18,"text":"$1d","time":1783597647,"title":"Ask HN: I hate coding agents. Is this skill issue?","type":"story"}],[48660884,{"by":"jjhartmann","descendants":3,"id":48660884,"kids":[48660995,48813008,48744275,48669403],"score":5,"text":"$1e","time":1782312764,"title":"Show HN: Sipp – Run small local LLMs in browser 3x faster","type":"story","url":"https://www.sipp.sh"}],[48806479,{"by":"getlawgdon","descendants":8,"id":48806479,"kids":[48806884,48807808,48829965,48830015,48877030,48809654,48837224,48812998,48829854,48810066],"score":3,"text":"$1f","time":1783353505,"title":"Ask HN: I use coding agents daily, but how do real engineers use them?","type":"story"}],[48810964,{"by":"achalpandey","descendants":9,"id":48810964,"kids":[48811785,48820377,48810999,48812705,48882087,48826789,48815116,48811231,48812989,48811258,48812971,48813197],"score":3,"text":"I think coding benchmark results don't represent our messy reality. As they

1) Use purpose-built test harnesses We use Claude Code or Codex

2) Test one-shot tasks We work in sessions

Sessions are messy, we start with a large primary task, then some cleanup, an adjacent fix here another over there. We start, stop, and change our minds.

That changes both cost and quality. Cache TTLs expire. Context grows.

I am working on creating one. Here's my rough plan: 1) Use Claude Code and Codex \n2) Use session shaped workloads. Stitch multiple SWE bench verified tasks into one big session. \n2.a) Use tasks from the same repo to ensure topical continuity. \n3) Headline metrics: dollar cost vs quality. \n3.a) Secondary metrics: turn count, time to completion

Open Questions \na) Does this problem make sense? \nb) Does my benchmark spec make sense? \nc) Are 10 SWE bench verified task starting with a hard one the right workload shape?","time":1783374736,"title":"How are you measuring Claude Code and Codex performance?","type":"story"}],[48751318,{"by":"mt_","descendants":3,"id":48751318,"kids":[48751347,48757111],"score":4,"time":1782930864,"title":"Amalia – an open-source language model targeting European Portuguese","type":"story","url":"https://huggingface.co/amalia-llm"}],[48748207,{"by":"sspehr","descendants":7,"id":48748207,"kids":[48758433,48749257,48749628,48756088,48764255,48750496,48756755,48757100],"score":4,"text":"Title basically, for me it feels like latest generations of LLMs are quite equal in usefulness for coding, does anybody have anecdotes of the opposite case?","time":1782918802,"title":"Ask HN: Does anybody still FEEL improvements between latest LLMs for coding?","type":"story"}],[48751300,{"by":"roschdal","descendants":3,"id":48751300,"kids":[48751799,48751364,48751459,48757088,48751770],"score":2,"time":1782930800,"title":"Ask HN: Will AI LLMs for coding get smarter and cheaper in the future?","type":"story"}],[48659077,{"by":"AndrewLiu96","descendants":1,"id":48659077,"kids":[48659841,48744263,48659096],"score":3,"time":1782306091,"title":"Show HN: An opinionated ranking of 21 open-weight LLMs, filterable by your GPU","type":"story","url":"https://northwoodsystems.ai/research/open-source-models-big-board"}],[48722052,{"by":"danboarder","descendants":53,"id":48722052,"kids":[48723122,48724416,48728717,48723119,48732119,48724280,48726582,48723479,48725316,48731067,48739990,48731984,48724527,48734619,48728573,48729307,48730966,48725092,48744250,48729052,48729768,48734470,48725384],"score":262,"time":1782753377,"title":"Ornith-1.0: self-improving open-source models for agentic coding","type":"story","url":"https://github.com/deepreinforce-ai/Ornith-1"}],[48721903,{"by":"stared","descendants":759,"id":48721903,"kids":[48723879,48722381,48722311,48722951,48724121,48726111,48724642,48729345,48722522,48722302,48724626,48725687,48724534,48722229,48724198,48725127,48724394,48722290,48722329,48722952,48727784,48722662,48722368,48727699,48722136,48722900,48723351,48724817,48729198,48725224,48723639,48722782,48724040,48722507,48722996,48723380,48723347,48722131,48725009,48725753,48722985,48732191,48725367,48724183,48722984,48729015,48725294,48728796,48724689,48727517,48724147,48722839,48726135,48722965,48730338,48726309,48725215,48722256,48726356,48725265,48728693,48727494,48729549,48730471,48725102,48729387,48723935,48724338,48723707,48722298,48730992,48725216,48726042,48732700,48726606,48722174,48728253,48725778,48724657,48725574,48727422,48728092,48727223,48746016,48732825,48724613,48726633,48730872,48734925,48724667,48729078,48729070,48728039,48733591,48728640,48722320,48729677,48734036,48731892,48728285,48723085,48724495,48729043,48729264,48730870,48725096,48730356,48728930,48733111,48729050,48724628,48726214,48726317,48722967,48727210,48726011,48722101],"score":1192,"time":1782752716,"title":"Qwen 3.6 27B is the sweet spot for local development","type":"story","url":"https://quesma.com/blog/qwen-36-is-awesome/"}]]

user:	modgate
created:	June 16, 2026
karma:	-1