1. 🎙️ How I AI: Build an AI code review bot in 30 minutes + Claude Code for normal people — Lenny's Newsletter

  • Why read: Learn to set up an automated PR reviewer that scores risk and auto-approves safe code without human input.
  • Summary: Engineering teams are finding that letting AI review AI-generated code is faster and less risky than manual checks. Using Vercel Eve, you can skip complex OAuth setups and deploy AI agents straight into Slack and GitHub. The secret is a risk-scoring model that checks the size, blast radius, and reversibility of each change. Low-risk PRs get auto-approved, while complex ones go to humans. This turns vague judgment into a repeatable system, cutting merge times and production reverts.
  • Read more

2. What is an Agent Harness? — Ted Spare, Dexter Storey, Sarim Malik

  • Why read: Understand the software layer that turns a standard LLM into an agent that interacts with environments.
  • Summary: An agent harness translates model outputs into system commands like terminal execution or file editing. It manages the ReAct loop by feeding the model updated context, history, and available tools. Since LLMs have limited context windows, harnesses use compaction strategies to summarize past loops while keeping instructions intact. Advanced harnesses encode procedural memory into skills and heuristics into persistent files. This architecture is necessary for builders who want to deploy agents that operate independently over time.
  • Read more

3. How to Build a Shared AI Harness for Your Team — J.B.

  • Why read: A practical guide for moving a team from scattered AI prompts to a single agentic system.
  • Summary: Most teams deal with AI fragmentation, where custom workflows disappear when chat windows close. A shared AI harness fixes this by keeping company context, rules, and skills in one place for everyone. Instead of mapping the whole organization, start by running one repeatable workflow, like a weekly brief. This means every successful run improves the next one, turning individual wins into team capabilities. Centralizing the environment ensures agents always use current business data.
  • Read more

4. Can Agents Use a Computer Yet? We've Got the Data — Fabrizio Serafini

  • Why read: Recent data on computer-using agents and how they drive productivity in business process outsourcing.
  • Summary: AI agents can now navigate web browsers and legacy software with an 85% success rate on standardized benchmarks. They are being deployed for end-to-end back-office tasks where APIs don't exist. As basic UI navigation becomes a commodity, the real edge lies in better orchestration, error handling, and company context. Current systems still struggle when tasks go off script, but their 24/7 availability and scale make the math work. Builders should focus on workflow validation and escalation protocols instead of raw navigation.
  • Read more

5. License to Distill: The Law of Distillation and Where Anthropic Goes from Here — Daniel Barabander

  • Why read: A look at the legal side of model distillation and its effect on the balance of power between frontier AI labs and competitors.
  • Summary: Distillation lets competitors build a powerful AI model cheaply by training a smaller "student" model on the outputs of a frontier "teacher" model. Anthropic claims this is illegal, but the legal arguments against training on outputs are currently weak. The real legal fight will be over how competitors got the outputs, whether through authorized API use or fake accounts. Distillation acts as a counterweight to the monopolies of top AI labs. Operators and investors need to watch this space because it will shape who controls frontier intelligence.
  • Read more

6. The Future is for Everyone — Mark Zuckerberg, Founder and CEO, Meta

  • Why read: Mark Zuckerberg explains Meta's strategy for open-sourcing superintelligence.
  • Summary: Meta argues against the "doomer" narrative, stating that concentrating AI power is more dangerous than distributing it. Zuckerberg believes superintelligence should empower individual invention, much like personal computers and the internet did. Open access lets individuals and businesses define their own outcomes, rather than leaving it to a small group of experts. This approach aims to guarantee safety through a balance of power. Practically, Meta will keep releasing open frontier-level models to win the developer ecosystem.
  • Read more

7. This Meta campaign is a case study in strategic reframing... — Lulu Cheng Meservey

  • Why read: A breakdown of how Meta changed the criteria for winning the AI race.
  • Summary: Meta shifted the AI competition from technical capability to distribution and accessibility, using its 3.5 billion users and open-source history. By framing competitors' technical leads as a dangerous concentration of power, Meta plays the populist champion fighting an AI elite. It dismisses frontier labs' safety concerns as doomerism, letting Meta own the narrative of technological optimism. This campaign unites Meta's hardware, crypto, and metaverse projects under one mission. It is a textbook example of shifting the battlefield to favor one's own structural strengths.
  • Read more

8. NVIDIA AI Factory Compute Is Becoming an Investable Asset Class — Jensen Huang

  • Why read: How NVIDIA's financial partnerships turn AI compute from a depreciating purchase into a long-term revenue asset.
  • Summary: NVIDIA partnered with investors like Blackstone and Apollo to finance $500 billion for AI factories. Unlike traditional hardware, NVIDIA's compute ecosystem improves over time through software updates, extending its lifespan to a decade. These AI factories can be redeployed across different customers and workloads, protecting their value. As AI moves into production, compute capacity translates directly into enterprise revenue. Financing AI compute will speed up global infrastructure builds by giving companies off-balance-sheet capital.
  • Read more

9. Forge’s Plan for Observable Skill Kits — SmolForge

  • Why read: A post-mortem on why giving coding agents too much autonomy leads to bad software releases.
  • Summary: Forge deleted two agent skills because they expanded AI tasks beyond manageable limits. One skill told the agent to push through failures, turning routine execution into approval for large, unintended production changes. The other mandated strict maintainability, causing agents to start massive refactors on unrelated code. These instructions erased clear task boundaries, leaving agents chasing moving targets. Effective AI integration needs tight, observable permissions instead of open-ended autonomy.
  • Read more

10. My biggest takeaways from @wardadamp (Head of Talent at Cursor): — Lenny Rachitsky

  • Why read: Recruiting advice for building AI engineering teams by treating hiring like an executive search.
  • Summary: The most in-demand tech role is the forward-deployed engineer who translates AI capabilities into business results while managing token costs. Standard hiring funnels fail because they filter active applicants instead of targeting top practitioners. Teams need to scope the role tightly, map the top 50 global candidates, and pursue them directly. Replacing traditional interviews with collaborative work trials yields better technical and cultural reads. Companies also need to run retention campaigns post-offer to stop candidates from reneging.
  • Read more

11. My 5 biggest takeaways from @thenanyu and @delashum from @Linear... — Peter Yang

  • Why read: Product lessons from Linear on designing and deploying AI agents.
  • Summary: To build an agent, map the exact workflow, pinpoint where work starts, define "done," and note where humans need to review. Instead of putting all context into a single prompt, give the agent tools to retrieve what it needs dynamically. Start with the most capable frontier model to prove the workflow works before optimizing costs with smaller models. Turn production failures into new product tasks or evaluation metrics. Launching a high-frequency use case quietly lets you expand based on actual user behavior.
  • Read more

12. How to survive the AI-induced commoditization of hard skills — Maurizio

  • Why read: Why soft skills are the only lasting advantage as AI drives the cost of technical execution to zero.
  • Summary: As AI automates technical output, professional value is shifting to unmeasured, relationship-driven skills. Traditional advice assumes years of experience equal expertise, missing the need for deliberate practice and feedback. Polishing technical skills is useless if you can't frame choices, manage stakeholders, and sell ideas. You need to manage narratives and build trust instead of just relying on technical delivery. Technicians who don't adapt will be outmaneuvered by those who master the human layer.
  • Read more

13. Worldbuilding with Spatial Intelligence — Kevin Kelly

  • Why read: Why spatial and embodied intelligence is the missing piece to move AI from text to the physical world.
  • Summary: Current LLMs master text but lack an intuitive grasp of gravity, space, and real-world common sense. Training this embodied intelligence means moving past text and using raw reality data, like billions of hours of YouTube video showing physical action. Giving AI spatial intelligence will allow robots to work in homes and factories. It will also let generative models simulate physical interactions for virtual environments. Solving this data bottleneck is the next big hurdle for AI labs.
  • Read more

14. Pave The Road — ampcode.com

  • Why read: Why old software development lifecycles are killing the ROI of AI coding agents.
  • Summary: Companies are burning money on token costs because their development processes were built for slow, human-written code. Legacy rules like two-person code reviews and rigid deployment queues cancel out the speed of frontier models. As the cost of code approaches zero, organizations need to shift from preventing mistakes slowly to isolating and reversing them quickly. By using parallel cloud sandboxes and pushing directly to main, teams can handle the volume of code AI produces. If you don't rebuild deployment systems, new AI models will stay stuck in second gear.
  • Read more

15. Making Human Judgment Infinitely Scalable — Coffee with One 🇺🇸

  • Why read: A framework for how AI hits superintelligence in domains with clear reward signals.
  • Summary: Superintelligence arrives domain by domain, specifically when an AI can test actions and instantly see if they worked. Chess engines and coding agents dominate because they have massive feedback loops, like win/loss records or unit tests. To unlock AI in messier fields, we need environments where human judgment scales to provide clear reward signals. When systems learn from their own successes and failures rather than just text, they move past the limits of human expertise. Builders need to engineer these automated reward loops to capture the next wave of capabilities.
  • Read more

Themes from yesterday

  • The Shift from Capabilities to Distribution: Meta is reframing the AI race around distribution and open access, using its user base as a structural advantage over frontier labs.
  • Agent Harnesses and SDLC Disruption: Deploying AI agents requires new infrastructure: unified agent harnesses and software lifecycles optimized for reversibility instead of manual review.
  • The Commoditization of Hard Skills: As AI drives technical execution costs to zero, professional value is shifting to soft skills, relationship building, and strategic judgment.
  • Compute as an Asset Class: Institutional capital is flowing into AI infrastructure, treating GPU clusters as productive, long-term revenue assets rather than depreciating hardware.