> ## Content Index > Fetch the complete content index at: https://www.antoinebuteau.com/llms.txt > Use this file to discover other available public pages before exploring further. # Daily Digest - 2026-07-29 - URL: https://www.antoinebuteau.com/daily-digest-2026-07-29/ - Published: 2026-07-30T08:04:40.000Z - Updated: 2026-07-30T08:04:40.000Z - Description: Explains why supervising an agent's step-by-step process is better for reliability than only checking its final output. Real-world tasks take time and involve many intermediate steps. - Author: Antoine Buteau - Tags: Digest **1\. Out of the box, long-horizon agents struggle to accurately perform... — X (formerly Twitter)** - Why read: Explains why supervising an agent's step-by-step process is better for reliability than only checking its final output. - Summary: Real-world tasks take time and involve many intermediate steps. If you only grade an agent on its final output, you can't tell if it followed a sound process or got lucky. The fix is to write "behavior specs" that outline the expected steps, like checking primary sources. Judges then grade the agent's full path against these specs, even if they don't know the exact final answer. This yields much better data on how the agent operates and ensures it follows your standards. - [Read more](https://twitter.com/mitch%5Ftroy/status/2082513195357307158/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **2\. this is how we built all of these sales/GTM workflows... — X (formerly Twitter)** - Why read: A practical guide to building in-house, automated sales workflows with AI agents. - Summary: Wiring Claude to tools like HubSpot and Slack lets you build custom sales workflows that beat generic off-the-shelf software. You can have agents study your closed-won deals and combine that with live data like recent hires or funding to pinpoint your ideal customer profile. Agents can figure out what works in your outbound campaigns and automatically generate lists of similar prospects. To protect email deliverability, set up and warm up alternate domains. Building these workflows yourself lets you constantly iterate and spot buying signals in real time. - [Read more](https://twitter.com/chrispisarski/status/2082236016161677644/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **3\. Lessons from Building a First-Pass AI PRD Reviewer at Uber — Uber Blog** - Why read: Shows how an AI reviewer can catch gaps in product requirement documents (PRDs) before they hit expensive human review meetings. - Summary: PRDs often show up to review meetings with missing context and ignored dependencies hidden deep in internal wikis. Uber built an AI PRD Evaluator to fix this. It scans drafts and pulls in related decks, past experiments, and metric definitions to check if a project is ready to launch. It acts as an early, fast reviewer for PMs, pointing out related work and past lessons. It doesn't replace human judgment. It catches the obvious gaps so teams bring much stronger, well-researched documents to formal reviews. Moving the AI upstream improves the work itself instead of overhauling the review process. - [Read more](https://www.uber.com/us/en/blog/first-pass-prd/?ref=antoinebuteau.com) **4\. The Code Is Disposable. The Codebase Is Not. — X (formerly Twitter)** - Why read: Argues that while AI writes code fast, it can slowly wreck a codebase's architecture if left unchecked. - Summary: AI agents write and review code so quickly that humans can't keep up with reviewing every line. Standard tests check if the code works today, but they don't catch long-term architectural rot. An AI change might pass all tests but increase coupling or make the code harder to read. Current benchmarks grade models on passing tests, ignoring long-term maintainability. Unless we find better ways to measure structural health, using agents for all coding and review will create codebases that run fine now but are impossible to update later. - [Read more](https://twitter.com/itamar%5Fmar/status/2082453557890945281/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **5\. CodeCrucible: A blueprint for LLM-driven SAST — block.xyz** - Why read: A guide to building AI security scanners that analyze entire repositories, outperforming older tools that only look at small snippets. - Summary: Building an LLM-based security scanner requires balancing context windows, finding vulnerabilities, and getting consistent outputs. Dumping a whole repository into one prompt often works well because the model sees the big picture, catching bugs that snippet-based tools miss. A two-step process works best: start with a broad search, then run a deep, specific analysis on what you find. This balances catching everything with minimizing false positives. To make it work in automated pipelines, you have to clean up and fix the model's inconsistent JSON outputs. The system's architecture matters much more than which specific model you use. - [Read more](https://engineering.block.xyz/blog/codecrucible-a-blueprint-for-llm-driven-sast?ref=antoinebuteau.com) **6\. Stripe and Sierra built their own coding agent systems. You probably should use Paperclip — X (formerly Twitter)** - Why read: Argues that engineering teams should use off-the-shelf platforms for running coding agents instead of building custom infrastructure. - Summary: Building a good coding agent system means managing sandboxes, handling permissions, and routing tasks. Tech giants build these in-house to merge thousands of PRs, but building the basics from scratch wastes engineering time. Open-source tools like Paperclip already handle task routing and permissions out of the box. Using an existing platform lets your team focus on writing specific skills and rules for your company instead of babysitting infrastructure. As agents do more work, you need a central system to control what they can access and how much they spend. - [Read more](https://twitter.com/dotta/status/2082522493483700525/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **7\. How well do frontier models use memory? — X (formerly Twitter)** - Why read: Shows how frontier models struggle to build and use memory effectively over time. - Summary: For agents to learn, they have to use existing memory and draw broad lessons from new experiences. Letta's new benchmark shows that while OpenAI models are cheap and good at using memory, they trail Anthropic in creating new memories. Weaker models just log specific user feedback instead of updating their instructions to handle future cases better. Over time, an agent's memory often degrades into duplicated rules and messy context, hurting performance. Testing models on long, real-world tasks reveals these memory failures that basic single-prompt tests miss. - [Read more](https://twitter.com/Letta%5FAI/status/2082375230690087129/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **8\. now that the dust is settling around the last wave... — X (formerly Twitter)** - Why read: A breakdown of the frontier model market, noting Grok 4.5's success and the problem with training agents purely on synthetic tasks. - Summary: Grok 4.5 is now a highly capable, fast daily driver, making SpaceXAI a clear peer to Anthropic and OpenAI. Its success shows the value of training models on real human usage data from popular consumer apps. In contrast, models like Opus 5 are struggling because training overly rewards passing long, automated tests while ignoring how well the model talks to humans. Fable 5 remains the smartest model available, but its high cost limits it to complex, high-value tasks. The market is splitting: users now have to route tasks to either the smartest models or the fast, cheap ones depending on the job. - [Read more](https://twitter.com/kunchenguid/status/2082525751606419830/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **9\. EVERYTHING IS AN EDGE CASE — X (formerly Twitter)** - Why read: DoorDash CEO Tony Xu on why applying AI to the messy physical world is so hard, and why it matters. - Summary: DoorDash focuses on AI for the physical world because an assistant that can't affect real outcomes isn't very useful. But the real world charges an "atoms tax": weather, traffic, and human error constantly break digital models. To get a return on AI investments here, companies must tie their AI spending directly to measurable customer outcomes instead of simply buying new capabilities. Also, since engineers only spend part of their day coding, AI has to speed up everything else, like product reviews and alignment meetings, to boost company-wide productivity. The goal is to take the productivity gains of your best individuals and upskill the whole company to match them. - [Read more](https://twitter.com/gokulr/status/2082488808000676156/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **10\. Why compute might get 10x+ more expensive in coming years — dwarkesh.com** - Why read: Explains why compute prices will likely spike as AI models get better at automating high-value work. - Summary: AI labs are growing fast, but to keep it up, they need much higher margins or much higher compute prices. Labs don't want to spend all their compute on inference because doing so implies they've stopped training newer, better models. When the next generation of models can automate highly paid work like senior software engineering, the dollar value of a GPU hour will jump. Labs and cloud providers will then charge a premium for large compute clusters. This shift will price out low-margin applications and smaller companies, leaving only the players who can wring the most revenue out of the hardware. - [Read more](https://www.dwarkesh.com/p/why-compute-might-get-10x-more-expensive?ref=antoinebuteau.com) **11\. The Multi-Model Future is Here — X (formerly Twitter)** - Why read: Argues the future belongs to an ecosystem of specialized models that beat massive, general-purpose models on specific tasks. - Summary: AI's "Bitter Lesson" favors general training methods, but that doesn't mean we need one giant model for everything. General language models are versatile but lose out on performance and efficiency for specific jobs like logical reasoning or deep search. The real goal is getting the most capability per unit of compute. Specialized models trained for specific goals, like AlphaZero for chess, can beat massive frontier models in their own domains while using a fraction of the compute. The future will rely on a mix of specialized models because they are faster, cheaper, and more accurate for narrow jobs. - [Read more](https://twitter.com/maxrumpf/status/2082520111756554442/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **12\. Microsoft Rents the Frontier — tomtunguz.com** - Why read: Contrasts Google and Microsoft's cloud strategies, showing the financial risks of renting AI models and chips versus owning them. - Summary: Google Cloud is growing faster than rivals because it owns its models and its custom TPU chips. This yields better margins and lets them expand capacity aggressively. Microsoft, however, runs on thinner margins: it rents Nvidia chips and relies on OpenAI's models. Almost half of Microsoft's future backlog is tied to OpenAI, a company that relies on venture capital instead of internal profits. This heavy reliance on one customer creates credit risk, forcing Microsoft to limit its capital spending despite high demand. Renting the frontier means giving away the biggest economic gains of the AI boom. - [Read more](https://read.readwise.io/read/01kyrh5rj2vywpk4gn90h71xpj?ref=antoinebuteau.com) **13\. AI billing infrastructure: Rethinking the build-versus-buy decision — stripe.com** - Why read: Explains why AI companies need flexible billing systems to survive rapid pricing changes without burning out their engineers. - Summary: Pricing AI is tricky and changes constantly; it drives product design and user behavior. Startups often guess wrong and either bleed cash on power users or fail to charge enough for the value they provide. Because AI moves so fast, building your own rigid billing system wastes engineering time that should go into the product. On the other hand, old SaaS billing tools can't handle the high-volume, metered usage that AI agents generate. Buying flexible, third-party billing software lets teams iterate on pricing quickly without maintaining complex infrastructure. - [Read more](https://stripe.com/en-ca/guides/ai-billing-infrastructure?ref=antoinebuteau.com) **14\. 0%, 50%, or 200% — danshapiro.com** - Why read: A blunt look at the choices leaders face with AI productivity gains: do nothing, cut staff, or double output. - Summary: When AI makes workers twice as effective, companies have three choices. The first is to do nothing, which creates "secret cyborgs" who use AI to work fewer hours while the company gains nothing. The second is to keep output flat and lay off half the staff to cut costs. The third and winning approach is to keep the team and use their new capacity to double output, build more products, and take market share. Companies that try to shrink their way to success will get crushed by rivals who use AI to grow. - [Read more](https://www.danshapiro.com/blog/2026/07/0-50-or-200/?ref=antoinebuteau.com) **15\. The AI cat is already out of the bag — Angular Ventures** - Why read: Argues that trying to regulate and contain AI is pointless now that powerful open-weight models are everywhere. - Summary: Leading AI labs want strict global licensing and export controls, but the spread of open-weight models is a done deal. Developers worldwide use distillation (training smaller, efficient models on the outputs of the giants) to match top-tier performance cheaply. The constant lobbying for regulation by massive AI labs suggests their actual technological moats are too shallow to protect their heavy investments. Since anyone can download a frontier model and run it locally, innovation is completely decentralized. The AI boom won't be a controlled rollout; it will be a messy, global explosion of productivity. - [Read more](https://link.mail.beehiiv.com/ss/c/u001.KDyGkvzFwiyqlBj6RxytbA1iPSAXtSZRozD5cC07tEVYFMsTMBS90ZYAz9FU7PzAL2YFUJlhbWiPv8EqnGBffPSF2nSbJATlhAgkFWtf%5FD-HuZS9ON53yAcxu6LU4tAvSAbpIZ%5F-4WU288Ux8OO%5FUdDgXfGLhlwXjYxcjAG3WqHQcl3KyKjrU1foqnbuHVGkhpbcgs-UpG%5FgBHeGsJAlcKJKekJfus8cGQiY2zyiRoneL87vJCFvyfgUb16760ZMJ7XkGRI11lWvToiSxA%5FuXPFkyD8bCwsx%5FBo2keGdrCHcIj7SWbueo0owbMxoVz2ijkWR1LQXkVLvKQgZ%5FCm7SRutuq0yDzgpPnB7AwWHFk41vKLpJrCSwHMfn6DTzwojIj0SDyDMNwKzPTbC0qYGOD7paV5ncRA3W7uGbkBP2oxlzVoqNbNkss4%5FCr5xzvn4BHgq%5FAeNqhi%5FLNRJj5ZNQwQZ77dnUtDYrffSpsLeP7fOJlX6Zb%5FEjJgFupjqaT-MNb88swRczOF3hRG0jQwHRX49izdgDdmLFDsmHOo6Vxh%5Fl5kB5v8dHNYBDtxufIQh82RwNz-rF4FPvB%5FQqmdD8YaKM0DJOKU8iFyQh1RFUjqk631jGLcnQBB1ae7t3h3kvz1Qi7Smw8gteIEvLHS%5F1DCsrj7r8lAIxxg2Rex2xu8/4sq/bpyDsAVrTIqQDvnKzzp1Cg/h0/h001.DyjSnmU2TOkMUq7NlP2rxqUn51qKlXENE1u3Hu3uZek?ref=antoinebuteau.com) ### Themes from yesterday - **Agents in the real world:** Companies are learning that complex tasks require more than just code generation. They are focusing on checking an agent's process, preserving code architecture, and using off-the-shelf platforms. - **The model market splits:** Powerful open-weight models and specialized tools are breaking the monopolies of the major AI labs, pushing the industry toward using multiple models for different jobs. - **The cost of intelligence:** Rising compute prices are forcing cloud providers and AI startups to rethink their margins and billing systems. - **What to do with productivity:** When AI makes workers twice as fast, executives have to decide whether to cut staff to save money or use the new capacity to grow the business.