1. Building for Trillions of Agents — Aaron Levie

  • Why read: A strategic roadmap for the shift from human-centric to agent-centric software design.
  • Summary: Levie argues we are crossing a chasm where agents (Claude Code, OpenClaw) move from basic chatbots to persistent entities with sandboxed compute and long-term memory. In a world with 1000x more agents than people, software must be redesigned to be "consumable" by AI rather than just humans. Agents will soon be the primary "users" making adoption decisions, meaning the most successful platforms will be those easiest for an agent to sign up for and navigate. This shift requires a "Paul Graham-style" refocus: instead of making what people want, start making what agents want.
  • Link: https://twitter.com/levie/status/2030714592238956960

2. Harness Engineering for Coding Agents — Dex

  • Why read: Technical blueprints for building reliable AI agents beyond simple prompting.
  • Summary: "Harness engineering" involves building the deterministic systems and environments around a model to prevent repeated mistakes. The post emphasizes that better context engineering always beats longer context windows; giving the model less to think about yields higher-quality results. Practical tips include using CLIs and `SKILL.md` files for progressive disclosure of information rather than overwhelming an agent with too many tools. The best harness engineers are minimalists who prioritize deterministic control via hooks and recursive sub-agents.
  • Link: https://www.humanlayer.dev/blog/skill-issue-harness-engineering-for-coding-agents

3. Now I’m The Bottleneck — Seiji

  • Why read: A visceral look at the "tipping point" where AI throughput exceeds human cognitive capacity.
  • Summary: With the release of high-speed parallel agents like Replit Agent 4, the traditional "request and wait" workflow has collapsed. The user experience has shifted from "ordering at a restaurant" to "working the line in a kitchen" where multiple agents are constantly waiting on the human for the next instruction. This creates a "vibe coding" paradox: AI makes work easier but forces the human to work significantly harder to maintain the flow of ideas. We have officially moved into an era where the human mind, not the compute or the code, is the primary system bottleneck.
  • Link: https://twitter.com/seijadvice/status/2032831857268764911

4. Self-Improving Skills for Agents — Vasilije

  • Why read: A technical proposal for moving from static prompt files to living, evolving agent capabilities.
  • Summary: Current agent skills (like `SKILL.md`) are static and break as environments change, but "cognee-skills" introduces a self-improvement loop. By storing skill execution data in a graph—including task patterns, errors, and user feedback—the system can reason about its own failures. When enough evidence of "weak outcomes" accumulates, the system can autonomously amend the skill's instructions or tool calls. This transforms skills from fixed prompt files into living system components that adapt to codebase shifts.
  • Link: https://twitter.com/tricalt/status/2032179887277060476

5. Yann LeCun’s AMI Lab Raises $1.03B — Chamath Palihapitiya

  • Why read: Details on the massive bet against LLM-only architectures in favor of "World Models."
  • Summary: Yann LeCun’s new lab, Advanced Machine Intelligence (AMI), secured a record $1.03B seed round to pursue Joint Embedding Predictive Architecture (JEPA). The core thesis is that AGI won't come from scaling next-token prediction, but from specialized "Superhuman Adaptable Intelligence" that understands causal reasoning and physical constraints. This move signals a post-LLM path for AI, focusing on architectural innovation for robotics and healthcare over raw compute scaling. Simultaneously, Netflix's $600M acquisition of InterPositive highlights the move toward bespoke, per-film AI models in creative industries.
  • Link: https://chamath.substack.com

6. Bimodal Hiring: The Middle is Death — Gokul Rajaram

  • Why read: A blunt assessment of how AI is hollowing out the middle-tier labor market.
  • Summary: The hiring market has split into two extremes: 10x "super-engineers" who refine their craft with AI, and young, "AI-maxxed" doers who are fearless in execution. The "middle"—characterized by spreadsheet jockeys and traditional middle managers—is facing obsolescence. Success in this new regime requires either extreme domain depth or extreme agent-driven agility. For operators, the practical implication is clear: avoid the "middle" by becoming either a high-leverage architect or a high-velocity AI pilot.
  • Link: https://twitter.com/gokulr/status/2032856787721007543

7. The Doc Read Process Has Killed Amazon — Bryan Beal

  • Why read: A cautionary tale of how once-innovative corporate rituals can become bottlenecks in the AI era.
  • Summary: Amazon’s famous "doc read" culture, while successful for decades, has slowed decision-making to a halt in a high-velocity AI world. The requirement for humans to sit and read physical documents in silence is inherently at odds with AI-native workflows that can summarize and synthesize instantly. While competitors iterate in real-time, Amazon's process forces multiple "rewrite and re-read" cycles. This serves as a warning for any organization: if your "sacred" processes prevent you from adopting AI speed, the process will eventually destroy the company.
  • Link: https://twitter.com/bryanrbeal/status/2032635436510675015

8. Stewardship Over Status: My Principles — Kaz Nejatian

  • Why read: Foundational leadership principles from Shopify’s COO on agency and systems.
  • Summary: Nejatian advocates for "stewardship over status," arguing that ownership forces the clarity and urgency required to leave things better than they were found. He emphasizes "Defaults Over Everything," noting that systems should be built so that the right action is the automatic one. At Shopify, the "Say the Thing" mantra prioritizes truth over comfort, which he views as a primary competitive advantage. For leaders, high standards are framed as a "gift" that signals belief in a team's potential.
  • Link: https://nejatian.com/principles

9. AI Code Review: The 100% Acceptance Future — Andrew Chen

  • Why read: A provocative prediction on the inevitable erosion of human-in-the-loop engineering.
  • Summary: A survey of founders shows a 50-50 split between those who review all LLM-generated code and those who just accept it. Chen predicts this will move to 100% acceptance as the only way to maintain the high throughput required of "AI-native" teams. Reviewing every line becomes a bottleneck that negates the speed gains of using agents. The practical implication is a shift in engineering from "line-by-line auditing" to "system-level validation" and automated testing.
  • Link: https://twitter.com/andrewchen/status/2032902478388449462

10. This is Not the End: The Jevons Paradox of AI — Unemployed Capital Allocator

  • Why read: A rational counter-argument to "white-collar doomerism" based on historical work expansion.
  • Summary: The author argues that white-collar work is subject to the Jevons Paradox: as work becomes cheaper and more efficient via AI, we don't do less of it—we do 10,000x more. Just as spreadsheets didn't kill accounting but exploded the volume of financial modeling, AI will enable humans to attempt "harder problems" that were previously cost-prohibitive. The "last 1-5%" of accuracy will become the new high-value battleground as volume scales. The future isn't a lack of work, but a shift toward managing the massive scale of AI-generated output.
  • Link: https://twitter.com/atelicinvest/status/2032925657550958852

Themes from yesterday

  • The Human Bottleneck: A recurring realization that agent throughput has officially surpassed human ability to "review" or "wait," requiring new delegation mentalities.
  • Agentic Architecture: A shift from "chatbots" to "harnesses" and "self-improving skills" as the primary unit of AI development.
  • Organizational Velocity vs. Ritual: The tension between traditional corporate decision-making (Amazon's docs) and the raw speed of AI-native competitors.
  • Post-LLM Strategy: Massive capital moving toward "World Models" (JEPA) and specialized AI, signaling a move beyond simple next-token prediction.