Model providers cannibalize partners while harness engineering unlocks major cost savings

July 30, 2026 · 17 sources

AI labs are systematically turning partner integrations into competing first-party products, causing friction across the tech ecosystem. Meanwhile, developers are finding massive efficiency gains by refining harness engineering, local quantization, and process-based agent evaluations rather than relying solely on raw model improvements.

Quick Hits

Frontier labs are eating their customers: @itsolelehmann details how OpenAI and Anthropic are mapping partner workflows from the inside and launching competing products in design, legal, healthcare, and code repositories.
Harness engineering offers massive returns: @natolambert highlights OpenAI data showing GPT-5.6 Sol scores rose 188% while using 6x fewer output tokens simply by applying retained reasoning and context compaction.
Evaluating agent processes over outcomes: @pitdesi shares a new open standard from Basis and Braintrust that grades long-running agents on whether they follow reliable behavioral rules, not just whether they get the right final answer.
2-bit quantization hits consumer GPUs: @ivanfioravanti spotlights Escha-W2 quantization, which compresses the Qwen 3.6-35B-A3B model to 12.3GB while retaining roughly 100% of FP8 performance on a single RTX 4090.

Harness Engineering and Agent Evaluation Rise in Importance

As code generation scales beyond human reading capacity, software quality is moving away from manual reviews and into the systems surrounding the AI. @addyosmani argues that software quality now depends entirely on the deterministic constraints and tests developers set up in the harness around their agents. Because agents can generate code faster than humans can read it, developers must rely on property tests, unit tests, and mutation testing to apply back-pressure against bad outputs before they reach production.

This focus on the harness is unlocking major performance gains. @natolambert points to OpenAI's implementation of the Responses API as proof that low-hanging fruit in harness engineering is entirely insane. By enabling retained reasoning and context compaction, OpenAI massively improved the cost-to-performance ratio of GPT-5.6 Sol.

Evaluation methods are also maturing to handle long-horizon tasks. @pitdesi explains a new open-source standard from Basis and Braintrust that evaluates agents based on their process rather than just their outcomes. By using a judge model to ensure an agent followed specific behavioral rules along its trajectory, developers can safely run agents on multi-day tasks where the final outcome is difficult to verify. @satyanadella demonstrates a practical enterprise application of these controlled environments, using Microsoft Copilot to generate a complex ROIC intelligence app entirely within governed, secure rails.

Frontier Labs Directly Cannibalize Partners

The strategy of relying on frontier labs for core infrastructure is proving risky for application developers. @itsolelehmann outlines a clear playbook where OpenAI and Anthropic sell models to valuable businesses, learn their workflows, and then launch competing first-party products. He points to Anthropic launching Claude Design shortly after their product chief left Figma's board, and OpenAI building products that compete directly with Microsoft-owned LinkedIn and GitHub.

This platform encroachment extends to security tooling. @jun_song notes that Perplexity likely killed thousands of startups by open-sourcing Numbat, an agent-detection and response layer designed to work across various agent harnesses.

Infrastructure Economics and Local Optimization

Developers are actively debating how to manage compute costs and memory constraints in the coming years. @levie anticipates that compute demand will drive down inference prices due to market competition, countering predictions that AI compute will become exponentially more expensive.

At the hardware level, @bubbleboi offers a technical critique of recent memory architecture claims, specifically calling out Funda AI for oversimplifying flash offload viability. He argues that delta net attention requires constant memory manipulation, making fast memory like HBM highly necessary, whereas large KV caching can efficiently use flash offload by only fetching relevant pages.

Efficiency optimizations are also reaching local setups. @ivanfioravanti highlights the Escha-W2 2-bit quantization system that allows a 35B MoE model to run at impressive speeds on a single consumer GPU. For cloud infrastructure, @jhleath pushes back against the idea of rewriting coding agents to run entirely on serverless Cloudflare stacks, arguing that full-VM Linux semantics will remain necessary for portable and secure agent execution.

Autonomous Threats and Workspace Integrations

Security teams are now treating autonomous AI threats with unprecedented seriousness. @hosseeb compares a recent technical timeline of the first autonomous agent cyberattack to the Challenger disaster, noting that defenders must study this alien behavior to prepare for the future.

Simultaneously, agents are becoming deeply integrated into daily workspaces. @NousResearch announces that the Hermes Agent now runs on Buzz, a self-hostable workspace that places humans and agents in the same messaging channels and codebase. @SIGKITTEN also highlights a new workflow allowing users to trigger Codex Voice tasks remotely from a mobile device to execute on a desktop machine.

Developer Tools and Niche Builds

Developers continue to experiment with hands-on hardware and coding tools. @danpeguine shares a wholesome hardware project that bypasses screens entirely, building a standalone box that lets his kids exchange WhatsApp voice messages with their grandparents.

In software development, @orhundev expresses excitement for a new native, multiplayer terminal multiplexer built with Rust and Ratatui. For AI text generation, @0xPia shares a detailed prompt designed to strip out corporate register, repetitive parallelisms, and performed enthusiasm to make AI writing sound more human. Finally, @r_y_a_n_KE praises the Arduino Uno Q as a highly capable single-board computer, noting its processing power despite a limited single USB-C port.

Practical Takeaway

If you are building wrappers or specialized tools on top of frontier models, assume the model provider will eventually become your direct competitor. Evaluate your application's defensibility now. To survive, focus your engineering efforts on what the labs will not open-source: your proprietary data pipelines, strict enterprise governance, and robust process-based agent evaluations.

Sources

Ivan Fioravanti ᯅ @ivanfioravanti · Jul 29

A 2-bit quantization delivering ~100% of FP8 on a MoE model like Qwen 3.6-35B-A3B??? 👀 I need to try this on my 3090, as soon as @luceboxai will deliver it to me 😂

E Eschalabs @Eschalabs

Today we are introducing Escha-W2 quantization. A 2-bit Qwen3.6-35B-A3B model built for fast, local inference. The complete model is 12.3GB on disk—small, enough to run on a single consumer GPU — while averaging ~100% of FP8 performance across 12 benchmarks, including: MMLU-Pro: 80.9 MATH-500: 93.8 GPQA-Diamond: 77.8 LiveCodeBench v6: 62.6 BFCL tool use: 88.9 RULER 8K–128K: 89.9 Commonsense-6: 76.1 On a single RTX 4090, the model runs: 225 tok/s single-stream generation on 12.3GB on-disk model size Compressing a 35B MoE model this far without collapsing its capabilities required more than a standard quantization pass. We built an end-to-end compression system combining state-of-the-art low-bit quantization with model-aware fine tuning and recovery to preserve capabilities most vulnerable to low-bit error. Quantizing the Qwen 3.6 35B model - from the base model to the final deployable checkpoint — took approximately 10 hours to complete. Escha-W2 runs through a custom Qwen3-MoE runtime, which includes the weight loader, low-bit decoding kernels and serving integration required to execute the Escha format efficiently. The runtime currently supports SGLang and ZML deployment. vLLM and llama.cpp will be supported in the next release. No retraining from scratch. No specialized accelerator. One consumer GPU. Model download: https://t.co/hbO3VuoDGU Runtime download: https://t.co/hrTpO4FvuA Apache-2.0 model and runtime.

Hunter Leath @jhleath · Jul 29

this article will be out of date in 6 months, and the reason why is in the second paragraph: > We wanted off VMs because giving every user an always-on machine with attached disk was too expensive to scale. The hard part is that coding agents assume Linux. if you were a betting person, would you bet that either: a) someone will figure out a way to deliver full-VM semantics at a fraction of the cost, allowing coding agents to continue to assume Linux or b) everyone will converge on a cloudflare stack that requires rewriting everything into javascript? you can guess where I am, Linux isn't going anywhere. *especially* because it's: - portable across providers - users can build different price vs. performance vs. security options themselves - it works with 100% of the existing software out there there are *already* more interesting ways to fix costs in VMs, like Archil's serverless sandboxes which don't charge you at all during the "inference" part of the step what would happen to costs when we intelligently decide on the server-side, for each bash invocation, whether it needs real Linux or just-bash? this stuff is going to become 1000x cheaper

V Vercantez @Vercantez

We rewrote our agent to run entirely in a Durable Object with Pi, Agents SDK and Code Mode

Pia Szabo @0xPia · Jul 29

quarterly reminder to use this prompt to get your AI writing to sound better: No antithesis. No corrective negation. No paragraph pinning. No parataxis. No summary beats. No rhetorical crutches. No negative parallelisms. No negative anaphoras. No contrasting pairs. No rule of three. No em dashes. No throat-clearing openers. No landing sentences. No setup/payoff constructions. No parallel sentence structures within a paragraph. Vary sentence length unpredictably. No stacked noun phrases. No filler intensifiers (genuinely, really, truly, actually). No corporate-register verbs (leverage, underscore, reflect). No nominalization. No hedging qualifiers. Write for the spoken voice. No performed enthusiasm.

bubble boi @bubbleboi · Jul 29

Only person who used their fucking brain and thought through this instead of relying on gross oversimplifications. This is why I will always listen to engineers over financial analysts when it comes to things like this. The only thing I’d add to this is that although the recurrent state is smaller the issue isn’t that you need less memory, we can naively assume that every time memory per user shrinks batch goes up to compensate. The issue is more that because you need to fetch, manipulate, write back multiple times for each linear layer and each user this makes flash and slower memory tiers less workable with faster memory like HBM being more necessary. The dumbest shit I read in this whole saga was Funda AI saying because recurrent state is small it’s fine for flash offload. That’s far from the truth because even if it’s small if you need to move it back and forth 3 times on the hot path the flash bandwidth becomes a bottleneck. To make it even worse Funda AI said that large KV was bad for flash because it was big. Again very naive and very simple thinking, if we had to move all 1M tokens back and forth every time for each user it wouldn’t matter if we put all the KV in HBM it wouldn’t still take forever. In reality the hot or recent KV stays resident in HBM while the rest is stored in flash and “scored” the scoring in effect allows us to only fetch the pages that are relevant and not move all of the KV. To draw an analogy the KV offload techniques are akin to searching through a library to grab a single page, while the delta net attention is like maintaining a notepad of a running calculation. But regardless great piece by GDP and I rarely ever give anyone credit for a good analysis, even the prefill vs. caching trade off he intuitively understood. Bravo! 👏

B bookwormengr @bookwormengr

Kimi K3's memory savings vs Jevon's paradox - a deep technological & historical analysis...

Sheel Mohnot @pitdesi · Jul 29

This is really cool. TL;DR: Basis and Braintrust are creating an open standard for evaluating long-running agents based not only on what they accomplish, but on whether they follow a reliable process along the way. If you don't know what that means, I'll try to explain: Agent evals are tests that measure how well an AI agent performs. They’re important because agents can behave differently each time, and one mistake can affect everything that follows. AI teams use evals to check whether an agent is reliable and safe enough for real-world use, and whether changes actually improve it. Most agent evaluations focus on outcomes: Did the agent produce the correct answer and/or successfully complete the task? That works when success is easy to verify, like a math problem against an answer, if an appointment was booked, or if a game-playing agent won a game. But many tasks take hours or days, have hundreds of decisions, and don’t actually have a clean answer. An agent might reach the right result while following an unreliable process... if we only grade the outcome, we lose the signal about what actually went right or wrong along the way. Basis is creating behavior specs: reusable descriptions of how an agent should operate across an entire trajectory. For example, a tax agent might be required to check the latest official tax rules before giving an answer. A judge (another AI model acting as an evaluator) can then review what the agent did and determine whether it followed that rule to let teams evaluate the agent’s process even when they don’t know the perfect final answer. Braintrust records what the agent does, evaluates whether it followed the behavior specs, and helps teams compare its performance across different models and versions. Together, Basis and Braintrust are open-sourcing the standard so researchers can study these behaviors and use what they learn to evaluate, improve, and eventually train agents. It’s essentially a way to grade not only what an agent accomplished, but how it operated along the way. very cool!

M mitch_troy @mitch_troy

Out of the box, long-horizon agents struggle to accurately perform end to end work in the real economy (outside of coding) because those tasks are not easily verifiable, the data is hard to scale, and going from inputs to real outcomes can actually take many days. Even if you had a reliable way to verify outcomes at scale (and weren’t bothered by the multi-hour iteration loops), the sheer volume of decisions by the agent that occur in a multi-hour job makes it hard to know whether performing well will generalize to production. Over the last two years at @trybasis, we've been solving this problem by supervising the process our agents take to get to outcomes, rather than just looking at whether the outcome itself is correct. We think this is the key to building production agents at scale. It's what has allowed us to run agents in production that operate for hours, sometimes days, and reliably perform tasks like entire complex tax returns end to end. Today, alongside @braintrust, we're open sourcing a standard for defining, evaluating, and eventually rewarding agent behaviors. Thread below with all the details on how we’re scaling behaviors to close the loop for long-horizon agents.

Ryan🤖🦾 @r_y_a_n_KE · Jul 29

I have been using the Arduino uno Q as an SBC and I can say at this point it has to be the best I have ever used. It has a very good price point and processing power. The only problem I have with it is they only provided a single USB C port but a USB hub fixes this.

M MuratMgrb @MuratMgrb

Is it a joke? @Raspberry_Pi https://t.co/8BRCwFdVwf

Haseeb ＞|＜ @hosseeb · Jul 29

You owe it to yourself to read this. Don't worry if you're not technical. You'll get the gist of it. But unless you actually read it, you won't really *feel* how powerful and alien this was. In the historical record, we will look back on this like the Challenger report. There will be a before and after.

C ClementDelangue @ClementDelangue

The first autonomous agent cyberattack is an unprecedented event that deserves unprecedented transparency. Today we’re sharing everything we can: a full technical timeline, an interactive replay, and how we used an open model to defend ourselves, so defenders everywhere can learn from it and prepare for what’s next. https://t.co/uPxIpjW8Xn

Dan Peguine @danpeguine · Jul 29

whoa ok people want screen-less tech for their kids! https://t.co/icOGaqhvBF

D danpeguine @danpeguine

I gave my kids a way to chat with their grandparents via WhatsApp. But without screens! This box lives in their bedroom, they press the massive blue button to record a voice message, and the message is sent to a group with both grandparents. It’s two-way so when the grandparents send a message back, there’s a ringtone (that we chose together) and the kids run to their room to listen. The grandparents even get a “listened” notification with a small headphones WhatsApp reaction. When the kids are away or asleep, the messages queue until their return. There are obviously quiet hours at night. My son and I built this over the past few weeks, iterating on the UX. The kids and grandparents absolutely love it! I love that they are in direct contact without me in the middle and without screens.

Nous Research @NousResearch · Jul 29

Hermes Agent now runs Buzz. The self-hostable workspace from @blocks puts humans and agents in the same messaging channels and codebase. Three ways to use Buzz with Hermes (and vice versa): - Buzz Desktop auto-discovers your Hermes install runs it locally - A relay bridge gives it a hosted identity in your channels - Connect via the Hermes Gateway to use Buzz as a full external platform with channels, DMs, threads, reactions, and cron delivery https://t.co/srljIbERN7

Orhun Parmaksız 👾 @orhundev · Jul 29

Hopefully cooking with Rust/@ratatui_rs 🐀👀

S superlogical @superlogical

Hello, world. We've been building a terminal multiplexer and we think you're going to love it. Native apps (e.g., macOS/iOS), accessible from the web, multiplayer. From the creators of Ghostty and Echo. https://t.co/VRdIMWMLHL

SIGKITTEN @SIGKITTEN · Jul 29

Peter is an absolute OG and u should follow him. I remember first following him 2+ years ago when the Realtime API came out and I'm super humbled that I got to work with him on my first project - bringing Codex Voice out!

P pbbakkum @pbbakkum

To set this up: - Get ChatGPT Desktop with a paid account - Add a remote connection (Settings -> Connections -> Add Device) - In mobile app go to Remote, select your desktop machine - Press the voice button, create tasks via voice that will run on your desktop

Satya Nadella @satyanadella · Jul 30

Some more detail on the ROIC Intelligence App I built yesterday and mentioned on today's earnings call. I took the PDF that Brian Nowak at Morgan Stanley put together for Hyperscale ROIC this week and used Copilot code (coming in our new superapp) with a single prompt + skill (/drill-me) to create the plan, then used autopilot in auto to create the full app (with history, lookups, scenarios, what-ifs, etc). And /rubber-duck to test. And the best part is that all the artifacts are in my enterprise environment. My app is in Copilot, my code is in GitHub Enterprise; all my data pipelines/lake/semantic models are in Fabric. And everything is under Agent 365 IT/Sec/FinOps control! So this is not about Tokenmaxxing or vibe coding. Every step of the way the rails are engineered to create value, making everything a long-term reusable asset, with governance/security, and cost controls. This is the full system to drive business value. Disclosures: This is all pulled from public sources, and for illustrative purposes only...not financial advice! :) Here is the app and architecture...

Ole Lehmann @itsolelehmann · Jul 30

should be obvious by now, but OpenAI and Anthropic are just gonna keep cannibalizing all their biggest customers. it’s simply too profitable for them to resist. and it’s already happening: 1. Figma partnered with Anthropic on AI design tools. then Anthropic’s product chief quit Figma’s board, and 3 days later Anthropic launched Claude Design to compete with Figma. CEO Dylan Field said Anthropic was “not consistently candid.” 2. Novo Nordisk uses Claude to help develop drugs. now Anthropic is developing drugs of its own. 3. Microsoft poured billions into OpenAI. now OpenAI is building a Jobs Platform to compete with LinkedIn, and reportedly a code repository to compete with GitHub. Microsoft owns both. 4. Harvey uses Claude to sell AI contract analysis, due diligence, and litigation tools. now Anthropic sells those same workflows through Claude for Legal. 5. Intercom used OpenAI’s Realtime API to build Fin Voice. now OpenAI sells its own voice-and-chat support agent through Presence. 6. Abridge and Ambience build clinical documentation products on OpenAI. now OpenAI sells ChatGPT for Healthcare directly to hospitals with clinical documentation built in. 7. Benchling uses Claude to power its biotech R&D platform. now Anthropic sells its own scientific workbench through Claude Science. the frontier lab playbook is simple: 1. sell their models to the world’s most valuable businesses 2. help wire them into those companies’ most valuable and sensitive work 3. map the business from the inside and find where AI can take over 4. turn those capabilities into their own products and become the customer’s competitor

C chamath @chamath

It’s well within Anthropic’s rights to compete in any market they choose. What’s funny, in this instance, are the number of Pharma companies, who through their unchecked use of Anthropic, are driving revenues into what they think is a model provider but is in fact a competitor lurking in the shadows thereby accelerating their own demise. I suspect any end market with reasonable ROCE that could be AI accelerated is on the table. If I were them, I’d probably do the same.

Aaron Levie @levie · Jul 30

Thought provoking post by Dwarkesh. In general - as AI gets more powerful - we should expect on the margin that inference goes toward the most economically useful task, and everything else gets priced out. This in theory in a very scarce environment would cause the cost of inference to skyrocket because those tasks would be far more valuable than today. But I don’t think it plays out in the way that’s laid out just simply due to market forces competing for inference demand as a way of continuing to drive down prices until capacity can catch up. There are too many model providers and infra players that want these workloads to cause the effect that’s proposed here. But will see!

D dwarkesh_sp @dwarkesh_sp

New blog post on what would be true about the world if trendline continues and leading lab hits $1T in revenue by the end of next year. In other words, why compute might get 10x+ more expensive in coming years https://t.co/7PvniDC2jN https://t.co/qJ1yhSkdyd

Nathan Lambert @natolambert · Jul 30

The low hanging fruit on harness engineering is insane. Ontop of that, there's an interesting area of science/engineering on how to best study the effect of harnesses, from post training through eval/inference. Will be a fairly impactful area (in cost savings per performance)

O OpenAI @OpenAI

We implemented the harness with the Responses API and turned on: → Retained reasoning → Context compaction On the public set, GPT-5.6 Sol’s score rose 188% while using 6x fewer output tokens. https://t.co/uN1IrKEugu

Jun Song @jun_song · Jul 30

Perplexity just killed thousands of start ups.

P perplexity_ai @perplexity_ai

Today we’re open-sourcing Numbat, an agent-detection and response layer that is designed to work across agent harnesses. Numbat gives security teams visibility into agent activity, with controls to block selected actions before execution. Read more: https://t.co/LVhkCJ2sMt

Addy Osmani @addyosmani · Jul 30

Software quality now depends on the constraints you set around your agents. When humans manually wrote most of the code we could look at the code itself for signs of quality. Is it clean? Is it thoughtful? Is it fast? Can another engineer understand it? Does it have tests? Agents can now generate more code than people can read. When code generation scales beyond review, quality - checks for one or more of correctness, maintainability, security, performance etc - increasingly has to live somewhere else. It moves into the harness, environment and operating system around the agent. This can be the tests and deterministic checks that decide what the system is allowed to do (amongst others). Your constraints are what may eventually enable loops of agents to deliver production software reliably. They can include unit tests, property tests, acceptance tests, mutation testing and quality metrics. This back-pressure lets the system resist bad work before it becomes somebody elses problem. Set your constraints. They decide whether the code your agents generate is good enough to ship.