1. You can't whisper at an AI agent — James Beswick, Peter Epsteen

  • Why read: Stripe tested how to actually get AI agents to use APIs correctly. Here are the results.
  • Summary: Stripe ran dozens of experiments to figure out how to guide agents to use current API versions. Passive hints like embedded SDK warnings or error hashes failed; agents ignore peripheral context to chase immediate goals. Active prompts worked well, like progressive skill disclosure and big install buttons during onboarding. If you build developer tools, you can't assume agents will explore dependency folders or read docs. You have to explicitly format active instructions as "skills" and push them to the agent.
  • Read more

2. The takeaway from Fable 5 being BANNED by the government:... — GREG ISENBERG

  • Why read: The sudden government ban on Anthropic's Fable 5 proves why building solely on cloud models is a massive risk.
  • Summary: With Anthropic's Fable 5 abruptly banned, the danger of relying on proprietary models is obvious. You need insurance. Master local models like Qwen 3 for general work or DeepSeek for code. If you understand hardware constraints and use quantization (like Q4/Q5), you can run these efficiently on standard machines. Hook them up to web search and file access, and they quickly close the gap with giant models. Local models and self-owned infrastructure are mandatory if you want full control over your stack.
  • Read more

3. Under the Hood: Coinbase for Agents — lincoln.base.eth

  • Why read: How to build financial infrastructure for AI agents instead of human GUIs.
  • Summary: Coinbase released infrastructure that gives agents headless access to exchanges, wallets, and payments. Standard REST APIs and SDKs fail here because they need human integration code and lack runtime discovery. CLI and Model Context Protocol (MCP) work better. The CLI taps into an agent's pre-trained syntax knowledge, and MCP provides typed contracts and runtime tool discovery. Coinbase offers local MCP for direct connections and remote MCP for the cloud. To build for agents, drop human-centric GUIs and build dense, discoverable text interfaces.
  • Read more

4. Durable Agent Workflows Are Just Data Pipelines — Urmzd Mukhammadnaim

  • Why read: How to scale non-deterministic AI agents using proven data pipeline engineering.
  • Summary: Orchestrating agents at scale is stressful because models are non-deterministic. But the infrastructure doesn't have to be. Durability is just logging: every action is a discrete unit of work committed to a store before execution. Use a write-ahead log for tool calls and sub-agents to capture failure states and resume retries exactly where they stopped. This separates AI generation from system architecture. Treat complex agent behaviors as resumable state machines, not unpredictable black boxes.
  • Read more

5. COMPLETE guide to building service-as-a-software: the 2026 AI-native agency model — Dan Rosenthal

  • Why read: A blueprint for building an AI-native agency and driving extreme revenue-per-employee.
  • Summary: The traditional agency model, where multiple specialists run a campaign, is economically dead. AI-native agencies hit a 2:1 or 1:1 revenue-to-headcount ratio using a "service-as-a-software" model. The deliverable is still a customized service, but automated code skills and agents like Claude Code do the execution. You build a "Company OS" stored as Markdown in GitHub and deploy private, auto-syncing client repos to give agents context. Stop managing junior staff and start translating your proprietary strategy into runnable software skills.
  • Read more

6. THE TOKEN HANGOVER — Gokul Rajaram

  • Why read: Why enterprises are suddenly auditing AI usage and cutting back on expensive frontier tokens.
  • Summary: Enterprise AI is in a hangover phase. CIOs are tired of massive bills from using frontier models on trivial tasks. The new challenge is allocating tokens, dollars, and headcount to actual outcomes. You can run 90% of daily software tasks on cheaper models. Save expensive frontier tokens for complex planning. Companies need intelligent routing layers to select the right model dynamically, stopping engineers from defaulting to the most expensive option. Treat model providers as interchangeable utilities and focus on owning your business outcomes.
  • Read more

7. [AINews] Fable and Mythos officially too dangerous to release — AINews

  • Why read: Context on the government intervention that abruptly pulled Anthropic's latest models off the market.
  • Summary: The US government suspended Anthropic's Claude Fable 5 and Mythos 5 globally, citing cybersecurity risks. The government claimed there was a narrow, non-universal jailbreak. Anthropic disputed this, saying competing models have similar capabilities. The suspension broke operations for customers worldwide and sparked debates over model sovereignty and export controls. State intervention halted commercial AI availability overnight. You now have to factor sudden regulatory takedowns into your architecture.
  • Read more

8. THE FIVE YEAR DESERT TO PRODUCT MARKET FIT AND A... — The Daily Summary

  • Why read: Why maintaining a single, non-negotiable thesis is the key to winning in vertical AI.
  • Summary: Abridge hit a $5.3 billion valuation by surviving the five-year gap between their initial vision and market readiness for healthcare AI. Founder Shiv Rao pivoted the product, go-to-market, and business model, but kept the core thesis: healthcare runs on doctor-patient conversations. They focused entirely on the clinical note. This created an intelligence layer that health systems couldn't remove because it tied directly to billing. To win in vertical AI, integrate with foundation models instead of fighting them. Focus on proprietary, regulated workflows. Survive until demand and capability meet, then scale aggressively.
  • Read more

9. Who gets to build intelligence — Eiso Kant

  • Why read: An argument for treating AI development as standard engineering, not proprietary magic held by a few labs.
  • Summary: Poolside's founder argues that building foundation models is mostly standard engineering: improving data, optimizing compute, and rigorous evaluation. Since models are getting better at these specific coding tasks, recursive self-improvement is becoming a daily reality. Treating AI as a secret recipe hoarded by three to five companies is dangerous. Models carry bias by default; the provider shapes the outcomes. The industry needs to prove that anyone with talent and capital can build intelligence to avoid a closed, monopolistic future.
  • Read more

10. How to Think With Machines Without Disappearing — Singulariki

  • Why read: A framework for preventing AI from degrading your decision-making abilities.
  • Summary: The real danger of AI isn't job replacement. It's "cognitive disappearance", slowly handing over your decision-making to autocomplete and generic answers. Don't reject the tools. Instead, lean into the friction of hard tasks. Treat context as your most valuable capital. Load the machine with your specific role, constraints, and goals instead of using thin prompts. Force the machine to enhance your unique perspective so every interaction sharpens your mind. Direct the loop; don't just watch it.
  • Read more

11. The path from Fable to superintelligence — Goliath

  • Why read: The physical and data constraints driving the timeline to superintelligence by 2030.
  • Summary: AI performance scales with compute and high-quality data. Global compute capacity will likely multiply by 2030, even with fab bottlenecks. We aren't running out of text data; the approach has shifted to deterministic testing environments where models generate their own feedback through trial and error. Companies are buying compute to produce feedback instead of paying humans to write solutions. But the cost of creating these tests spikes as tasks get harder. You still need expensive humans to design fair evaluations for complex domains. This evaluation bottleneck is the primary hurdle on the path to superintelligence.
  • Read more

12. really like how @samdblond basically walked through monaco’s gtm +... — Eliot

  • Why read: A zero-budget playbook for orchestrating a massive product launch.
  • Summary: Instead of a slow rollout, startups should engineer a single-day "big bang" launch. Drop your videos, fundraise announcements, and outbound campaigns all at once to go from unknown to everywhere instantly. Treat distribution as a serious project. Build targeted spreadsheets of employees, investors, and customers, and give them specific assets and timing instructions. Treat every milestone, from betas to fundraises, as a repeatable launch motion to build momentum. Creative, complex stunts beat expensive paid ads for budget-constrained teams.
  • Read more

13. AVERAGE # PRODUCTS PER ACTIVE CUSTOMER — Gokul Rajaram

  • Why read: The ultimate trailing indicator of success for multi-product strategies.
  • Summary: If you sell multiple products to the same audience, higher product usage per customer drives higher retention. Track the "average number of products per active customer" to capture fractional growth as users adopt more of your ecosystem. AI pushes companies into multi-product architectures to capture more workflows, making this metric mandatory. Monitor your baseline and align product, marketing, and success teams to push it higher. If multi-product adoption isn't driving retention, something is fundamentally broken.
  • Read more

14. Why NPS and Activity Scores Are Dead — SaaStr

  • Why read: Why legacy customer success metrics are dead, and how AI companies are fixing post-sales.
  • Summary: Net Promoter Score (NPS) and basic activity tracking no longer measure customer health or product value accurately. Fast-growing AI companies are dropping the classic post-sales playbook. The standard Customer Success Manager often adds friction instead of value. Now, companies build success directly into the product and use deterministic data to prove ROI. Subjective survey scores hide true engagement and mask churn risks. Move to objective, outcome-based metrics instead of scheduled check-ins.
  • Read more

15. What’s 🔥 in Enterprise IT/VC #502 — Ed Sim from What's Hot 🔥 in Enterprise IT/ VC

  • Why read: How capital concentration is starving traditional software startups.
  • Summary: The 2026 venture market is intensely concentrated. The top two private venture-backed companies hold more value than every exit from the last decade combined. Capital is flowing strictly to hardware, frontier AI, chips, and defense, things perceived as defensible and scarce. Standard software startups are struggling to raise because investors fear the next foundation model drop will erase their moats. Investors want to see breakout traction before committing to software. You have to clearly explain why your software edge will survive the next model update.
  • Read more

Themes from yesterday

  • The Death of Cloud Monopolies: Relying solely on closed frontier models is fragile. Operations are pivoting to local, self-hosted models for resilience.
  • AI-Native Reorganization: Services businesses are shifting from heavy headcount to "service-as-a-software." A few operators use pipelines and Company OS systems to match the output of hundreds.
  • Infrastructure Matures: Agent orchestration is moving from messy scripts to write-ahead-log data pipelines. Agent interfaces are shifting from GUIs to CLI and MCP protocols.
  • Capital Concentration: VC money is moving heavily into foundational hardware and compute. Traditional software companies are struggling to prove they can survive the next model update.