> ## 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 - 2026-06-17 - URL: https://www.antoinebuteau.com/daily-digest-2026-06-17/ - Published: 2026-06-18T10:45:54.000Z - Updated: 2026-07-19T00:09:33.000Z - Description: Agents work better when systems shape information and interfaces around their needs: filter command output at the source, move large data through files, and expose capable APIs and CLIs. Human judgment must still govern architecture and tradeoffs. - Author: Antoine Buteau - Tags: Digest **1\. RTK, Headroom, and the right way to save tokens in coding agents — Trevin Chow** - Why read: How to save tokens in coding agents by shaping command outputs instead of relying on generic compression. - Summary: Generic compression tools like Headroom struggle in coding agents because unpacking compressed data creates extra steps. RTK (Rust Token Killer) takes a different approach: it intercepts terminal commands and filters the output based on built-in rules for commands like \`git diff\` or \`cargo test\`. Sending only the useful parts of the output to the agent saves tokens without breaking the workflow. The takeaway is to densify data at the source instead of compressing it after the fact. - [Read more](https://twitter.com/trevin/status/2067375373449605381/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **2\. Filesystem-Pilling Your Vertical Agent — Peter Wang** - Why read: A look at using bash and the filesystem to simplify complex agent tasks like data enrichment. - Summary: "Filesystem-pilling" gives agents access to a real shell and filesystem so they can route data through disk files instead of their context window. For data-heavy tasks, agents can cap bash outputs and save large JSON files to disk. They then use tools like \`jq\` to pull out only what they need, keeping token use low. Passing instructions via text files also helps subagents break down and run tasks in parallel. - [Read more](https://twitter.com/BrainsAndTennis/status/2067345406699393176/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **3\. Designing for Agents — David Cramer** - Why read: Why companies need to treat AI agents as primary users and design API-first interfaces. - Summary: AI agents are changing how we interact with software, often bypassing web UIs entirely. Companies need to build APIs and CLIs designed specifically for machines, ensuring these interfaces match the quality of their graphical UIs. If an agent cannot navigate a product's API, that tool will get dropped from developer workflows. The future requires designing for agents first. - [Read more](https://twitter.com/zeeg/status/2067386142375514354/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **4\. Using Agents, Keeping Agency. — Devin Dickerson** - Why read: A warning about letting AI coding agents make architectural choices based on mainstream popularity. - Summary: Developers frequently let AI agents pick their frameworks and databases without considering the tradeoffs. Models skew heavily toward mainstream tools, often suggesting a popular choice like Python when a specific task requires Rust. This bias, along with manipulation from the generative engine optimization (GEO) industry, can lead to technical debt and security risks. Engineers need to treat AI recommendations with skepticism and retain control over architecture. - [Read more](https://devindickerson.dev/posts/using-agents-keeping-agency/?ref=antoinebuteau.com) **5\. stop telling Claude Code/Codex "do this" — Vox** - Why read: Prompt templates to make AI coding models act like senior engineers rather than junior developers. - Summary: Telling an AI coding agent to "fix this bug" limits it to basic tasks. Better templates prompt the AI to analyze requirements, map out edge cases, and plan the architecture before it writes code. The templates in this thread cover building features end-to-end, refactoring unfamiliar code, and optimizing performance. Setting these expectations upfront forces the model to produce better work. - [Read more](https://twitter.com/Voxyz%5Fai/status/2067237707483337118/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **6\. Shape Suffixes — Good Coding Style — Noam Shazeer** - Why read: A naming convention for tensor variables that makes machine learning code easier to read. - Summary: Standard variable names hide the tensor shapes needed to understand neural network code. Adding a shape suffix, like \`input\_token\_id\_BL\` for a tensor with Batch and Length dimensions, makes the structure obvious. Teams should pick a single-letter dimension key and stick to it across the codebase. This works in PyTorch, JAX, or any tensor library, making it much easier to spot dimension mismatches. - [Read more](https://medium.com/@NoamShazeer/shape-suffixes-good-coding-style-f836e72e24fd?ref=antoinebuteau.com) **7\. Book Review: Inference Engineering — Joe Barrow** - Why read: A review of Philip Kiely's guide to scaling machine learning models in production. - Summary: "Inference Engineering" helps teams avoid bad architectural choices when scaling ML models. The book covers everything from picking serving frameworks like vLLM and TensorRT to scaling infrastructure. The reviewer suggests skimming it first, then diving into specific chapters like quantization and speculative decoding when those problems arise. The models change fast, but the foundational principles here will last. - [Read more](https://twitter.com/barrowjoseph/status/2067239202060747215/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **8\. 643: SpaceX Goes Shopping ($60B for Cursor, What Next?), Bring Your Own Tokens Might Save SaaS, Shotwell's 0.1%, O… — Liberty’s Highlights** - Why read: Why spending money on high-quality tools pays off by removing friction from daily tasks. - Summary: People tolerate mediocre tools too often, both in software and physical tasks. Switching to a high-quality option, like a well-machined ratcheting screwdriver, turns a frustrating chore into something satisfying. Better tools reduce the mental drain of poor performance. Find the bad tools you currently put up with and replace them. - [Read more](https://read.readwise.io/read/01kvb5hvgm312ppz7qk6hgmgtd?ref=antoinebuteau.com) **9\. The sales compensation problem every AI company will eventually hit: How to pay sales on usage — Meghan Gill** - Why read: How AI companies moving to consumption-based pricing should restructure their sales commissions. - Summary: Traditional SaaS commission plans fail when AI companies charge per token or API call. In a consumption model, closing the deal is only the first step; revenue depends entirely on what the customer actually uses. Sales leaders have to choose whether to pay commissions based on committed spend, incremental usage, or gross usage, and each has tradeoffs. Tying pay to actual usage keeps sales teams aligned with real revenue. - [Read more](https://meghangill.substack.com/p/the-sales-compensation-problem-every) **10\. How to build your AI GTM system — Kyle Poyar** - Why read: The hidden backend complexity of managing global payments and self-serve billing. - Summary: Moving to self-serve billing introduces massive backend complexity. What looks like a simple checkout flow quickly turns into dealing with global tax compliance, fraud, and local payment methods. Using basic payment gateways often creates administrative headaches and compliance risk. Handing this off to a merchant of record takes care of chargebacks and regional taxes, making global expansion much easier. - [Read more](https://link.mail.beehiiv.com/ss/c/u001.akMajagfhhFb3UGKEEczEFkLTeDkEfm6YqJAiL7-OxEQQArsUzegMO-vvYkch%5FKpsPfldcF8et5sBGZYYTwfXedCNstcWdnQyzKfBnpu67YNIw0O7dJObNkIYITK1J4piP9bHN0zNN29-cmrZ0cVtkqK-0Itme2tUWi--KxmQuQDZbCh0W2GSeQIEncogl5IjJHFAoxpTIdWuoOpk7slxqK26kKFPHJbMmFaiTMTTDUhb-VSBllyUGY5MJILnd8QM1XASz5hqmE17W-K0Zb1J0-YIGVCwZII3jY4Yi6R0ZCYlkELedpeXDGE%5FPMMmZhUL87tICyfCmBJeUDeoqNXR-Oo8yyPSQFnAZPtGR4qcr%5Fw9Cjgu1I5AjvcR1rbIYnpkEqUDNSlZAqogwEarwjzOephVfZontvy8Bc2vxOpiWx0QvYY3sP-Srq2fQJZMIc5uQClQr7nwmMkIh8siqLRb1X5NpW65F7R19QVR46XRX6EYV8eUNXpKQBLLv%5Fh6OGfhYvuDUp0uKz3h-y5-TdHHrR7ld15EkGE8-aV2Dbbs042FwsdVCEuOfC50f4uSJnOoHz6NsHb7Bgiba46KJVBzEG9HqVQMO-fc4TqqTqtobp--K0ng24ANdcVXOwBuCSW0JyFnDqJ7Ceg9hpFd5z7P26%5Fyy-GcV6DbiQHiYPg4t7ymEZGrFEFcDz3gsDaBcID%5FddafZ-9GR0mtmqDuuawrFYnyUclYjJhPvkG9aTr5h7VOnUOXEy-I2K4WJohb%5F2JEPUN9qk-zAj42B%5FVMROpbA/4rk/2NduMa9WThCwAGtAL1jL7A/h1/h001.gaSpw2fz0aWPerIXgl1qKqNde3vT3CcNeTNSSUODSRw?ref=antoinebuteau.com) **11\. The AI Adoption Trap — Gavin Owens** - Why read: Why organizations need to build proprietary "token capital" to stay competitive as AI advances. - Summary: Depending entirely on external AI models puts your competitive edge at risk. Companies need to build "token capital": the internal knowledge, context, and workflows that accumulate as AI systems operate inside the business. This requires making models swappable so you do not lose institutional knowledge if you change providers. The workflows themselves become the real intellectual property. Keep control over your architecture and data. - [Read more](https://twitter.com/gavinowensnet/status/2067299874538717693/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **12\. Everyone Has an AI Thesis. This Report Has the Data — Emergence Capital** - Why read: Data on how enterprises are actually adopting AI tools and allocating their budgets. - Summary: Venture firms have plenty of AI theories, but this report provides the hard data. It shows how enterprises are spending their budgets and adopting new tools. The most successful AI startups integrate directly into existing workflows instead of asking customers to migrate to entirely new platforms. Evaluation metrics are also changing, focusing more on gross margins and retention than just user growth. - [Read more](https://read.readwise.io/read/01kvaz204vxjrejk1w8dre6zj7?ref=antoinebuteau.com) **13\. The basics of being filesystem-pilled — Peter Wang** - Why read: How forcing agents to route their work through the filesystem saves tokens and improves control. - Summary: Giving agents a real filesystem and shell lets them route work through disk files instead of inline data. This cuts down token usage on both inputs and outputs. Bash outputs can be capped and saved to disk, and large tool outputs can be stored as files so the agent only has to read the paths and JSON structure, not the entire payload. It changes how agents handle large context windows and run parallel tasks. - [Read more](https://twitter.com/BrainsAndTennis/status/2067345406699393176/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) **14\. The Three Eras of Software Pricing — Meghan Gill** - Why read: How the history of software pricing explains AI's move to consumption-based billing. - Summary: Software pricing has gone through three phases. Perpetual licenses required large upfront payments and created vendor lock-in. SaaS moved to subscriptions, but still required upfront commitments. The consumption era, started by AWS and now used by AI startups, bills customers based on actual usage. This shift requires completely new approaches to sales and go-to-market strategies. - [Read more](https://meghangill.substack.com/p/the-sales-compensation-problem-every) **15\. Parallel and End-to-End Goals for Big Tasks — Vox** - Why read: A strategy for using multiple AI agents in parallel to handle large software projects. - Summary: A single AI agent often stalls out on large codebases or complex features. A better method is to have the primary agent break the goal into independent pieces and spawn parallel subagents to handle them. Each subagent gets a clear deliverable. The main agent tracks progress, resolves conflicts, and merges the results. This keeps the system moving on massive tasks without getting stuck. - [Read more](https://twitter.com/Voxyz%5Fai/status/2067237707483337118/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com) ### Themes from yesterday - **The Shift to AI Autonomy**: Agents are moving from simple code assistants to autonomous systems that handle end-to-end architecture and manage parallel subagents. - **Adapting Business Models**: The move to consumption billing and API-first tools is forcing companies to rewrite their pricing, sales compensation, and go-to-market strategies. - **Rethinking Data Pipelines**: Techniques like "filesystem-pilling" and command-aware filtering are changing how agents handle external data to avoid hitting context limits.