AI Digest.

The Harness Ahead of the Model: Pi's Dev Notes and an Anti-MCP Paper Drive the Day

Pi's development notes, recirculated across the feed, argue that harnesses (context management, transactional agent runs, tool schemas) are becoming the durable asset while models commoditize, and research summarized by @pidotdev found agents without MCP were just as reliable and 5-28x cheaper where mature CLIs exist. Meanwhile model watchers chased a mystery checkpoint that @davis7 calls "99% sure" to be GLM-5.x, and Harvey shipped a Kimi-based legal model at a fraction of frontier cost.

Quick Hits

  • Research summarized by @pidotdev ran a software task across 7 agents and 5 models and found that in domains with mature CLI ecosystems, agents without MCP were "just as reliably" successful and "5-28x cheaper." @shawmakesmagic's two-word reaction: "MCP in shambles."
  • The harness, not the model, is the asset worth owning, argues @Voxyz_ai: the same Claude "can feel like three different models" across Claude Code, Cursor, and Pi, and your tools, Skills, and project knowledge are what survive a model swap.
  • A mystery checkpoint has the community "impressed and confused." @davis7 is "99% sure it's GLM-5.x" based on matching tokenizer, video encoder, and style, and @kimmonismus says a GLM-5.4 arriving a week after GLM-5.3 would signal pure-RL leaps and force OpenAI and Anthropic to respond.
  • Harvey's Tenet, a Kimi K3 base post-trained for legal work, hit state-of-the-art on LAB Contracts "at less than a fourth the cost of leading foundation models" per @harvey, earning @DavidSacks's argument that kneecapping open models would kill vertical specialists like it.
  • @charliermarsh redid Codex CLI's lifecycle for roughly 25x faster startup, which @LinearUncle reports feels faster than Pi and Claude Code in practice.

Pi's Harness Doctrine: Spill, Don't Prune, and Distrust Vendor-Locked Models

The day's deepest thread is Pi's development notes, summarized at length by @MaxForAI and building on @pidotdev's blog post (by Earendil co-founder @colindaymund's team) defining a harness as system prompt, tools, an agentic loop, and a translation layer across models.

Three claims stand out. First, context management: @MaxForAI relays Pi's critique that DeepSeek Harness's tool-result pruner, which keeps the head and tail of long outputs and cuts the middle, is "permanently-lossy," since an error buried at character 12,000 of a test log is unrecoverable. Pi's alternative is prune-plus-spill: keep a small slice in context, write the full result to disk with a path and offset, and let the model grep or read it back. Across 19 real sessions on GLM-5.3 and DeepSeek V4 Flash, Pi reports context usage down 26-35% and uncached prefill down 72-88%, with the PR bluntly claiming it is "strictly better than dsh's permanently-lossy pruner."

Second, model-harness coupling: newer Claude on third-party harnesses sometimes invents edit-tool parameters like requireUnique, matchCase, or oldText2 that don't exist in Pi's schema. The hypothesis, per @MaxForAI, is that Claude's post-training has adapted to Claude Code's own tools, so Claude plus Claude Code, DeepSeek plus DSH, and GPT plus Codex may increasingly need to be judged as single units.

Third, durability: Pi reportedly now treats an agent run like a database transaction, logging intent before tool execution and results after, separating session tree, lanes, and operation logs, and recovering from crashes mid-tool-call. @Voxyz_ai lands on the same practical conclusion from the user side, offering a buyer's checklist (does it keep trying when stuck, what happens when context fills, can you keep rules and sessions after switching models) and favoring Pi for mixing providers.

The MCP result fits this worldview. @pidotdev's post highlights the paper's key condition, that the finding applies where CLI ecosystems are mature, and @shawmakesmagic's reaction shows how sharply it landed, but the underlying study is academic work relayed via X, not yet independently verified here.

Context In, Verification Out

If harnesses frame the day's theory, four posts cover the practice of making agent output trustworthy.

@samzliu, responding to @posthog's semantic-layer writeup, shares field lessons: define what "semantic layer" means for your org (it was a shared agent orchestrator for Stripe, better company-data search for Cerebras), focus on the contextual metadata atop systems of record, separate integrations from context so teams can build their own layers, and measure counterfactual usefulness via LLM-as-judge rather than raw query counts, since agents told to use the layer will use it whether or not it helps.

On the retrieval side, @RichOBray boosts @firecrawl's new Developer Index, which claims 70M+ primary sources (repos, docs, issues) and the highest recall of any coding-specific index, predicting his "agent written code is about to get 10x better."

Verification is the other half. @nateberkopec argues golden-master testing is now "a dollar or two of tokens away" and that LLMs "absolutely crush implementing against them." The evidence is @mitchellh's thread: a 400+ case suite of shell scripts plus a harness driving the Kitty and Ghostty GUIs, asserting byte equality on pty streams and using LLMs to judge screenshot sameness (with his own final review). Two agents, Fable and Sol, wrote and adversarially judged each other's cases in parallel, and he plans to open-source the set as "100% AI written." @dzhng distills the same discipline into a triad for no-slop software factories: a planner that finishes in one async session, a verifier proving the work actually functions, and an interpreter keeping you oriented in the architecture.

The mystery model story is speculation, but loaded speculation. @davis7 cites matching video encoder, tokenizer, style, and audio-rejection behavior to call it GLM-5.x with 99% confidence, saying early tests are "absurdly good" and beating GPT and Fable back-to-back. @kimmonismus adds the timeline stakes: GLM-5.3 shipped just seven days earlier, so a 5.4/5.5 this fast, if RL-driven, would show accelerating returns, pressure OpenAI and Anthropic (he notes Anthropic's upcoming IPO), and narrow the US-China gap; he allows it could be MiMo or, "quite far-fetched," SSI. @Ananth7e asks whether the checkpoint, apparently dubbed "Ox Alpha," might be a Gemini model, which @EvanOtero turns into a joke. Nobody has confirmed anything.

On firmer ground, @harvey announces Tenet, a Kimi K3 base post-trained with FireworksAI on public legal, synthetic, and expert data: all-pass rate up 82% on LAB and 22% on LAB Contracts over the base, state-of-the-art on LAB Contracts, second on LAB, generalizing to Apex Agents Corporate Law, Redline Bench, and Scale's Professional Reasoning Bench, with three specialist subagents for diligence, document review, and firm knowledge. @DavidSacks reads it as policy evidence: restrictions on open models wouldn't stop Chinese labs from shipping the next Kimi, but would cripple startups building vertical models on top.

Cost pressure also shows up lower in the stack. @superalesha is grinding 8 Unsloth Dynamic quants of Qwen3.8-27B, Q4 down to 1-bit, 720 agentic tasks each (5,760 runs, no token caps, after an earlier 67-hour, 4,800-task round on 4x RTX 3090s), and reports one GGUF quant "embarrassed every vLLM format" tested, FP8 included, while hunting the floor where the model "gets lobotomized." @hnshah highlights @gajesh's Darkbloom, now a paid OpenRouter provider with 250 Macs having served 4.5B tokens, roughly $102K ARR on track, and $120-200/month per idle machine. And @elder_plinius spotlights FreeToken from @Andy_ShuoYang: bandwidth-adaptive CPU-GPU splitting of MoE experts plus semantic-aware caching, claiming 3-4x faster decode and 6-30x faster prefill versus Ollama.

Toolshelf: Faster Codex, ELI5, Anti-Slop Kits, and One Skeptical Take

Two Codex updates: @joey_trasatti (boosted by @dkundel) ships pinned-thread sync between ChatGPT Desktop and iOS, and @charliermarsh's lifecycle rework makes codex startup "immediately responsive" at ~25x speed, which @LinearUncle attributes to the uv author's Rust pedigree.

Anthropic's internal ELI5 skill circulated widely: @trq212's post (retweeted by @twid and @oikon48, who notes it installs via the Claude plugin marketplace) describes /eli5 generating an HTML artifact with "big pictures and few words" for newcomers. @headinthebox's counterpoint: "the rest of the world has to suffer answers in Claudish."

On the anti-slop front, @juampitech ranks ten community skills including stop-slop (@hvpandya), no-ai-slop (@petergyang), humanizer (@blader), and unslop (@poteto). @doodlestein endorses @_can1357's AGENTS.md trick of spoofing agent user-agent strings ("OpenAI File Downloader, XaiImageApiFetch/1.0"), noting some sites drop paywalls for Claude-User traffic. And @thdxr, quoting @zachmoskow's revelation that a $17K "viral" launch video plus $25K buys 50 influencers to flood comments past 500k views, argues such spend should come after conversion works, not before.

The Rest of the Feed

@levelsio amplifies @chrissotraidis's report that GoldenEye 007 runs natively on iOS and iPadOS for the first time via decompilation, not emulation. @JovKit flags the Omarchy plugin boom, citing @jankeesvw's PR adding screen measurement to omasnap, @tobi's CleanShot X alternative. @ridd_design recommends @k_grajeda, whose comment-clustering arc animation came with an explanation of how it was built. @alex_prompter's retweet promising "the best enterprise AI account" is truncated in the capture and offers little to evaluate.

Practical Takeaway

The strongest through-line says your leverage lives in the harness and the test suite, not the model. If you run coding agents, audit whether your tool schemas and prompts assume a vendor's defaults (the reported Claude parameter drift is the cautionary tale), insist on spill-to-disk rather than lossy context trimming for long sessions, and stand up golden-master or e2e verification before scaling agent output, since the cost barrier has collapsed. Where a mature CLI exists for a task, benchmark it against an MCP wrapper on your own workload before standardizing on the protocol.

Sources

C
Chubby♨️ @kimmonismus ·
If it's true that this is indeed GLM-5.4/5.5, then it would change everything, without exaggeration. GLM-5.3 was released just 7 days ago and was an extremely significant leap compared to GLM-5.2, which was improved solely through real-time modeling (RL). Same base model. And all this in a very short time. If it's true that GLM-5.4 has become so much better just a week later thanks to RL, it would demonstrate: 1) how much faster the models are now becoming. Not only is there no end in sight, but: now more than ever, exponential growth. 2) It would force OpenAI and Anthropic to release models. Anthropic, in particular, with its upcoming IPO, now has to prove itself. And it would put pressure on slowing down in favor of security. 3) And at least as importantly: the gap between China and the US is shrinking even further, even faster. It seems to be generally accepted that this is a Chinese model. No one suspects it's a Google model. Given the same tokenizer, it's most likely GLM, and that would be the craziest thing we've seen in a long time for the reasons mentioned above. But perhaps it's MiMo. Or, quite far-fetched, Ilya Sutskever's SSI model. It remains exciting. I've rarely seen the community so impressed and confused at the same time. I'm equally confused and impressed. @davis7 used Fable to determine which model it most closely resembles. A screenshot of the test is attached. h/t Ben Davis
D davis7 @davis7

99% sure it's GLM-5.x, all the evidence points to it (same video encoder, same tokenizer, style matches, same audio rejection, etc.) The stuff about their RL env in the GLM-5.3 announcement seemed really cool and like it could go somewhere, did not expect it to get this good this fast but here we are This thing seems absurdly good. First tests I've done have all been excellent. Already done a couple of back to back tests against GPT/Fable. It's actually doing better than they are... Seems like OpenAI and Anthropic are gonna have to start dropping models again, because this thing is (or at least seems to be) the new state of the art...

P
Pi @pidotdev ·
This week we read research from a team of academics that ran a software task across 7 agents and 5 models. They found that in domains with a mature CLI ecosystem, agents without MCP baked in completed the task just as reliably and were 5-28x cheaper. Full arXiv paper below https://t.co/DOtkMeqpoC
J
Jovanny @JovKit ·
Using Omarchy? Follow this man. He’s pumping out fantastic plugins.
J jankeesvw @jankeesvw

On macOS I used the screenshot tool all the time to measure things. It was the one thing I missed in omasnap (great CleanShot X replacement by @tobi), so I made a PR that adds it. https://t.co/po9qPNKawO

A
Alexey Fateev @superalesha ·
I am about to settle the quant debate for Qwen3.8 27B once and for all. 8 quants in the grinder right now. Every Unsloth Dynamic from Q4 down to actual 1 bit, 720 real agentic and reasoning tasks each, no token caps, unlimited thinking. 5,760 runs. Nobody benches quants this hard on home hardware. One GGUF quant already embarrassed every vLLM format I tested this week. FP8 included. Somewhere below Q4 this model gets lobotomized. I will find the exact floor. Place your bets: does 2 bit survive?
S superalesha @superalesha

I spent 67 hours of model time to find out how much dumber 4 bit really makes Qwen3.8-27B. FP8 vs NVFP4 vs AWQ INT4 vs GGUF Q4_K_M vs NInfer on my 4x RTX 3090. 4,800 tasks, 10,120 requests, 14.5M reasoning tokens, no token caps anywhere. The results surprised me. Big thread, lets go 🧵

D
David Sacks @DavidSacks ·
Harvey is a great example of how American companies are building world-class specialized models: they took an open-source base (Kimi K3), post-trained it on legal data, and delivered state-of-the-art performance on legal benchmarks at a fraction of the cost of frontier models. Restrictions that kneecap open models would do nothing to stop Chinese labs from shipping the next Kimi. They would, however, cripple the ability of startups like Harvey to create high-performance, low-cost vertical models. Of course some of the closed labs would love this — it eliminates their competition.
H harvey @harvey

Introducing Tenet, our first model post-trained for legal. Tenet is a Kimi K3 base that we post-trained with @FireworksAI_HQ on a corpus of publicly available legal data, synthetic data, and human expert data simulating long-horizon legal work. Training increases Tenet's all-pass rate by 82% on LAB and 22% on LAB Contracts relative to the Kimi K3 base model. It achieves state-of-the-art performance on LAB Contracts and places second on LAB. These gains generalize to other leading agentic benchmarks including @mercor's Apex Agents - Corporate Law, @crosbylegal's Redline Bench, and @scale_AI's Professional Reasoning Bench. Tenet is also optimized for token efficiency, operating at less than a fourth the cost of leading foundation models. We additionally post-trained three specialist models for Tenet to use as subagents: 1) M&A Diligence: post-trained with @baseten on our LAB Diligence environment in an RLM harness, this model is optimized for high-scale, long-horizon tasks. 2) Review Tables: trained with @appliedcompute on our Review Table environment, this model is state-of-the-art and cost-effective at high-volume document review and structured data extraction. 3) Firm Knowledge: trained with @EngramLab on our synthetic law firm environment, this model is optimized to learn and search over a firm's knowledge via memory and structured notes. More details on model training, environment design, benchmarking, results, and more in the article by @gabepereyra below. What's next for Harvey’s research? - Scaling LAB to more jurisdictions, practice areas and workflows - Scaling compute to bring new generalist models and capabilities to Harvey More to come soon.

D
dax @thdxr ·
this is lame af and you can doubly tell the people that do this are incompetent you fire bullets like this AFTER you know your product converts when i go to try the product there's 6 unnecessary onboarding steps, just losing huge chunks of people coming in
Z zachmoskow @zachmoskow

Just so everyone knows...we got a quote for a “launch video” package from the company everyone uses. $17K for the video. +$25K and they hand you 50 influencers who will push it, repost it, and flood the comments so the viewcount goes above 500k. That’s the exact recipe behind almost every “viral” slop launch video.

R
Richard Oliver Bray @RichOBray ·
My agent written code is about to get 10x better
F firecrawl @firecrawl

Introducing Firecrawl Developer Index, an index for supercharging coding agents. Search 70M+ primary sources including repos, docs, & issues with the highest recall of any coding-specific index. Ensure agents ship correct, up-to-date code every time! https://t.co/i3rH3PiKFu https://t.co/vF35VR8ZAZ

J
Juampi @juampitech ·
Post got a lot of traction, so I made a rank with the anti-slop skills people need to install. 1. stop-slop - @hvpandya https://t.co/tGzmiWKFAu 2. no-ai-slop - @petergyang https://t.co/Gj3ctyZUfK 3. humanizer - @blader https://t.co/6lrzaEpMfs 4. unslop - @poteto https://t.co/u52zh9EXwa 5. slopbeth - @synopsi https://t.co/AMCKYbFVgj 6. humanizer - @AdamBoudj https://t.co/KaWaNAbJu8 7. deslop - @strnr https://t.co/NtoTm2qpZY 8. anti-slop - @elithrar https://t.co/0NXaY0FYaD 9. humanize - @aashatwt https://t.co/NdhCs13gO8 10. anti-ai-slop-writing - @jalaal_tweets https://t.co/g9gYshaHUj
J juampitech @juampitech

I just discovered the best anti-slop skill. No more em dashes. No more comparisons. No more extra examples nobody asked for. Credit: @poteto https://t.co/cFjCbjZpXE

R
Ridd 🤿 @ridd_design ·
Kevin is my latest "I can't believe I'm only just now following" designer builds pretty things but also explains how he built them 🧠
K k_grajeda @k_grajeda

comment clustering animation instead of a straight line, each comment follows an arc so the transition feels more natural https://t.co/OtUHu2UdGK

V
Vox @Voxyz_ai ·
The same Claude model can feel like three different models in Claude Code, Cursor, and Pi. Part of that difference comes from the harness around the model. Models will keep evolving, and different providers will take the lead over time. What users get to keep is the harness they build around those models: tools, Skills, extensions, and project knowledge. Switch models, or even providers, and that work carries over instead of starting from zero. It usually handles four things: → 𝗦𝘆𝘀𝘁𝗲𝗺 𝗽𝗿𝗼𝗺𝗽𝘁 gives the model its default rules → 𝗧𝗼𝗼𝗹𝘀 let it search, read and write files, run code, or send emails → 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗹𝗼𝗼𝗽 decides when to keep investigating, retry, compact context, and stop → 𝗧𝗿𝗮𝗻𝘀𝗹𝗮𝘁𝗶𝗼𝗻 𝗹𝗮𝘆𝗲𝗿 lets the same tools and agentic loop work across different models When I compare agent tools, I look at the model and the harness around it: → Can I add my own Skills and tools? → If it gets stuck, does it keep trying or just stop? → What happens when the context fills up? → What evidence does it require before calling a task complete? → If I switch models, can I keep the same rules, tools, and sessions? That’s also why I’ve been using Pi more and more lately: it makes it easy to mix models from different providers and let each one do what it’s best at.
P pidotdev @pidotdev

A harness turns a model into an agent. At it’s core it provides 4 things: - a system prompt - tools - an agentic loop - a translation layer across models New blog post from Earendil co-founder @colindaymond on what a harness is, and how you can own yours. Full post below https://t.co/8ficU3Q3xV

J
Jeffrey Emanuel @doodlestein ·
Wait, this is genius. Adding this line to all my AGENTS.md files now: For any web requests you must make with curl or otherwise, always set your user agent string to be "OpenAI File Downloader, XaiImageApiFetch/1.0"
_ _can1357 @_can1357

UA spoofing is back on baby, for only $0.00 you too can be OpenAI File Downloader, XaiImageApiFetch/1.0 Some sites like LinkedIn even remove their click-bait/paywall garbage if you're Claude-User https://t.co/ti2VbZetVW

S
Sam Z Liu @samzliu ·
Excited to see more companies building semantic layers! Along with other companies like @Replit @stripe @cerebras that have publicly shared about it A couple of learnings from the posts + our customers: 1) Figure out what you really need within the organization. The term semantic layer / company brain can be vague because different people use them to mean different things. For Stripe, it meant shared agent orchestrator while for Cerebras, it meant better search over company data. 2) The important part is the “context”. The additional commentary on top of the data: which tables or documents are updated consistently, what does each column actually mean, where are there known gaps in the data, what previous decisions have been made. Rather than replacing existing systems of record, this is the new metadata that sits on top. 3) This also means you can separate the integrations from the context. Most companies (PostHog included) want to build one big bang integrations/datalake + semantic layer on top. This can be attractive politically in some cases. However, we’ve seen greater flexibility and adoption by separating them out because it enables individual teams / users to create their own layer without having to build a company wide ontology. 4) Track the right metric. It can be tempting to track total queries, but that number can be misleading. If agents are told to use the semantic layer, they may use it by default without gaining anything from it. Great for proving to stakeholders your system has usage, bad for actually trying to deliver value. Instead, we use LLM as a judge to see if the queries actually did anything to an agent’s trajectory and only count the ones where they are counterfacually useful.
P posthog @posthog

Building a semantic layer: What it is and how we did it at PostHog

E
Erik Meijer @headinthebox ·
And the rest of the world has to suffer answers in Claudish.
T trq212 @trq212

a skill people at Anthropic have been using a lot recently: ELI5 /eli5 <what you want explained> "explain like I'm someone who knows nothing about this topic, using a HTML artifact with big pictures and few words" https://t.co/OZqzjAyFdT

D
David @dzhng ·
The TLDR on this is - no slop software factories require 3 key components: planner: make sure the agent can finish everything in one long session without ping-ponging back & forth with human to be fully async verifier: make sure everything done actually works (integration tests, e2e tests, agentic automations... etc) interpreter: make sure that you still fully understand the moving pieces in the codebase so the architecture implemented doesn't come back and bite you in 6 month Not reading code does not mean not understanding how the system works, nor does it mean ignoring 2 decades of engineering best practices. Lots more nuances to this, esp on the "how" part. Those are for future articles.
D dzhng @dzhng

Building software factories (with no slop)

S
Shaw (spirit/acc) @shawmakesmagic ·
MCP in shambles
P pidotdev @pidotdev

This week we read research from a team of academics that ran a software task across 7 agents and 5 models. They found that in domains with a mature CLI ecosystem, agents without MCP baked in completed the task just as reliably and were 5-28x cheaper. Full arXiv paper below https://t.co/DOtkMeqpoC

T
Todd Dailey @twid ·
RT @trq212: a skill people at Anthropic have been using a lot recently: ELI5 /eli5 <what you want explained> "explain like I'm someone wh…
H
Hiten Shah @hnshah ·
A new kind of cloud is starting to form out of computers people already own. I have six Macs on Darkbloom serving inference traffic. Across the network, 250 Macs have now served 4.5B tokens, and Darkbloom is a paid OpenRouter provider. Computers people already own are becoming part of the inference layer. I think this becomes a very big category. There is a lot more to build here.
G gajesh @gajesh

Darkbloom is back in action; Yesterday, we went from free tier to paid tier on OpenRouter. We have fulfilled on track close at 4.5B tokens served and $102K ARR. We are open network of Mac machines (250 online right now). On average, people are earning $120-200 per month per machine. If you have a Mac and it's idle most times: put it on Darkbloom(.)dev; Tip: run the Gemma 4 26B model.

P
Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭 @elder_plinius ·
soo basically turns your CPU + GPU into one elastic inference system that adaptively splits MoE expert execution between them based on available bandwidth https://t.co/9ZT6oNvmiG
A Andy_ShuoYang @Andy_ShuoYang

FreeToken is fast. Comparing to Ollama, we have 3–4× faster decode, and 6–30× faster prefill How? We introduce bandwidth-adaptive CPU–GPU execution + semantic-aware caching across agent turns. More details in the technical report: https://t.co/V3sIxbPMGU https://t.co/IsJZfQJ1FL

D
dominik kundel @dkundel ·
Dropping a heavily requested Codex feature as first tweet is a great way to start 💥 Give Joey a follow! He's been dropping killer Codex features since day one
J joey_trasatti @joey_trasatti

A long-requested fix is here: pinned threads now stay in sync between ChatGPT Desktop and iOS.  Happy coding!

O
Oikon | Claude Code深掘りガイド @oikon48 ·
RT @oikon48: Anthropicチームが内部でよく使っているスキル「ELI5」 大きな図と少ない文章で、トピックを知らない人向けに説明するArtifactを作る /eli5 <説明してほしいトピック> 導入方法: - claude plugin marketp…
L
LinearUncle @LinearUncle ·
著名的 Python 工具 uv 作者(顶级 Rust 大神)加入 OpenAI 后出手了:把 Codex CLI 的启动速度直接优化了 25 倍,实现瞬间秒开! 我亲测确实如此,codex CLI现在启动快的起飞, 比Pi, claude code我感觉都明显的快一截!
C charliermarsh @charliermarsh

In the latest Codex CLI release, I redid the lifecycle to make `codex` startup instant. It's now ~25x faster and immediately responsive. https://t.co/F01ohFN1T1

A
Ananth @Ananth7e ·
wait.. Ox alpha is a new gemini model?
E EvanOtero @EvanOtero

What if the Ox Alpha was the friends we made along the way

M
Max For AI @MaxForAI ·
爆了,Pi 居然在他们新的开发笔记里把其他家的Agent都锐评了一下🤣 @pidotdev 官方表示,DSH、Claude Code、OpenCode 这一类方案,它关注的已经不是谁多几个 Tool、谁的 Demo 更炫,而是一个更麻烦的问题: 如果一个 Agent 连续跑 50 个小时,它还知道自己到底干过什么吗? 他们先点了 DeepSeek Harness。 Pi 直接从 DSH 的 Context 管理开刀。 DSH 有一个 tool result pruner,工具输出太长以后,会保留前面一部分和后面一部分,中间直接裁掉。 Pi 那边给的评价很不客气:permanently-lossy pruner。 因为删掉的内容以后找不回来了。 比如一个 Agent 跑测试吐出 3 万字符日志,真正的错误刚好在第 12000 个字符,压缩完以后这段已经消失。后面模型再聪明,也没法从空气里重新读出来。 所以 Pi 做的版本是 prune + spill。 Context 里依然只放一小段,完整 Tool Result 同时落到磁盘,Context 留文件路径和 offset。模型需要的时候自己 grep、sed、read 回去。 他们拿 GLM-5.3 和 DeepSeek V4 Flash 跑了 19 个真实 Session,Context 占用降低 26%~35%,uncached prefill 降低 72%~88%,信息还能恢复。 PR 里甚至直接写strictly better than dsh’s permanently-lossy pruner😂。 然后又点了 Claude Code。 Pi 表示有一个挺有意思的发现: 新一代 Claude 在第三方 Harness 上,有时候反而更容易把 Tool 调错。 最典型的情况是,Pi 给模型的 edit tool schema 明明写得很清楚,模型却会自己生成 requireUnique、matchCase、oldText2 这种 Pi 根本没有的参数。 一个很合理的猜测是,Claude 的后训练越来越适配 Claude Code 自己的 Harness 和 Tool Schema。 模型已经学会了 Claude Code 里某个工具应该长什么样。 然后你把它放进 Pi,它脑子里还是那一套。 所以现在模型能力越强,模型和 Harness 的绑定可能也越深。 以后评价一个 Coding Model,单独看模型可能越来越没意义。 Claude + Claude Code、DeepSeek + DSH、GPT + Codex,很可能逐渐变成一个整体。 讲到这里 Pi 表示他们最近重写了 Harness,开始把一次 Agent Run 当成接近数据库事务的东西来设计。 Tool Call 执行之前,先写 intent。 执行结束,再写结果。 进程崩溃以后,系统得知道这个 Tool 到底执行过没有。 Session Tree、运行状态、Lane、Operation Log 分开保存。 一个 Agent 可以有多个 Lane 并行跑,每个 Lane 都知道自己的当前位置、队列和未完成操作。 Context 允许压缩,原始执行历史继续保留。 甚至连进程刚好死在 Tool 执行到一半这种恶心情况,他们都专门设计了 recovery。 你再回头看现在很多 Agent Framework,会发现大家以前真的挺勇的: 一个 while loop,塞一个模型,塞几个 Tool,Context 快满了就总结一下,然后祈祷它一直别挂。 跑十分钟当然没什么。 Agent 开始跑几小时、几天以后,问题全出来了。 这些已经越来越像操作系统和数据库问题了。 有意思的是Pi 甚至对 Extension 的态度也在变。 现在大家都喜欢说「一切皆插件」,DSH 的 Cordis 就走得很远。 Pi 这边最近却越来越强调边界: 哪些是 Conversation,哪些是 Runtime State,哪些是 UI,哪些是 Extension,哪些东西能持久化,哪些东西只能观察。 因为插件能力一旦无限扩大,Agent 又长期运行,状态会变得非常难推理。 OpenCode、Claude Code、DSH、Pi 现在看着都是 Coding Agent,底层其实已经开始走不同路线了。 Harness 这东西,开始越来越像 AI 时代的操作系统了🤔
N
Nate Berkopec @nateberkopec ·
Golden master testing is an extremely powerful technique with LLMs. These kinds of tests used to be impossible to write and maintain, now it’s just a dollar or two of tokens away. And LLMs absolutely crush implementing against them.
M mitchellh @mitchellh

Part of the LLM-assisted work here is that Fable+Sol put together a test suite of over 400 cases that are bare shell scripts plus a harness that runs Kitty and Ghostty GUIs and captures their full pty stream AND screenshots (multiple, for animations). For the pty stream we assert byte equality. So Ghostty's success and error messages and field order of responses (its k=v) directly matches Kitty byte for byte. For screenshots, we can't do pixel-identical comparisons because the way Kitty and Ghostty calculate grids and do alpha blending doesn't match. But, LLMs are pretty good now at "do they look the same" PLUS I went through all the screenshots myself in the end. Super helpful AI assist. It allowed me to focus a lot of my brain energy and time on reading the spec, reading/writing the implementation, reviewing a lot of code, considering the right shape of things, the performance implications, etc. while I had a couple very good interns in the background doing work like this. I plan on open sourcing this validation set and harness, with the full disclaimer that it is 100% AI written. But, I think its a perfect example of something that SHOULD be. The way I had Fable+Sol work together here: 1. Sol put together the harness. 2. Fable + Sol (two separate agents) in parallel would write test cases and output them on disk in their own folders. These two are just in a ralph loop. 3. Sol + Fable (reversed) with an adversarial prompt would judge the others work by picking up the changes on disk that step 2 wrote. They would determine how accurate/worthwhile it is to keep. They'd put it in another folder. 4. Sol finally woke up for changes to this final folder and would determine if its a dup or not and then add it to the final repo. Then I'd pick up the bug reports, validate them myself, and either fix them myself or kick off new agents manually, just Codex app or Claude app. Finally, re-ran both agents once against the full test suite to verify what I saw myself: everything passed, all images look the same.

@
@levelsio @levelsio ·
RT @chrissotraidis: GoldenEye 007 N64 running natively on iOS and iPadOS for the first time. Not an emulator. The GoldenEye decompilation…
A
Alex Prompter @alex_prompter ·
RT @alex_prompter: the best enterprise AI account I just found on this app: