1. The Anatomy of an Agent Harness — Vivek Trivedy

  • Why read: Explains how to build the systems that surround LLMs to get actual work done.
  • Summary: An agent is just a model wrapped in a harness. If you aren't building the model, you're building the harness. The harness provides what models lack: state, tool execution, and environment setup. Key parts include system prompts, infrastructure, orchestration logic, and hooks. Working backward from the behavior you want lets you extend and correct models. This shifts focus from model intelligence to the engineering required to make it useful.
  • Read more

2. A new and possibly controversial perspective: — Richard Sutton

  • Why read: Richard Sutton explains why generative AI struggles to produce outputs that are both novel and high-quality.
  • Summary: Generative AI uses supervised learning to mimic examples. It can generate outputs that are novel (via randomness) or good (via training data), but rarely both. When systems push beyond their source material to create something new, it usually reads as a hallucination. This mimicry is useful for making processes faster and cheaper, but it prevents the AI from making genuine scientific discoveries. Deploy generative AI for mimicry rather than autonomous innovation.
  • Read more

3. 🥇Top AI Papers of the Week — DAIR.AI

  • Why read: Three papers on optimizing AI agents for production workflows.
  • Summary: Recent research shifts focus from training larger models to optimizing surrounding systems. Microsoft's SkillOpt treats natural-language instructions as trainable parameters, using validation gates to refine them for better performance. Another paper shows how to compile agentic workflows into the weights of smaller models, cutting inference costs by 100x while keeping quality high. A third introduces AutoScientists, a decentralized team of agents for research. The most cost-effective approach for operators right now is optimizing the harness and orchestrator rather than relying on raw model capability.
  • Read more

4. some observations on costs of frontier ai: model performance is... — Apoorva Mehta

  • Why read: A look at the plateauing performance of frontier models and the economics of enterprise AI.
  • Summary: Performance gains between new frontier models are shrinking, forcing companies to scrutinize the ROI of upgrading. For many tasks, older or open-source models work fine when paired with fine-tuning and better harness engineering. The cost of running frontier models at scale drives companies to find cheaper alternatives. Operators should optimize data and tooling around cheaper models instead of defaulting to frontier options.
  • Read more

5. Anthropic Raises $65B at $965B, Passing OpenAI — Chamath Palihapitiya

  • Why read: A look at the capital war in AI and the emergence of decentralized model training.
  • Summary: Anthropic raised a $65 billion Series H at a $965 billion valuation, passing OpenAI in private markets, and launched Claude Opus 4.8. Meanwhile, alternative approaches to heavy capital expenditure are appearing. Pluralis Research trained a 7.5B model across 198 cities using consumer GPUs via Protocol Learning. This highlights two tracks of AI development: centralized, heavily funded frontier labs and decentralized computing networks. Operators should track both. High valuations demand massive enterprise adoption, but decentralized protocols could democratize future model training.
  • Read more

6. the solution might be cancelling my AI subscription — hmmz.org

  • Why read: How AI tools amplify distraction and lead to unmaintainable code.
  • Summary: AI tools let developers prototype rapidly, but this often leaves a graveyard of unmaintained repositories. Generating code with zero friction acts as an ADHD amplifier, encouraging endless side quests instead of focused problem-solving. This promotes token usage over deliberate engineering. Friction often forces necessary focus. Churning out code serves no one if the products lack purpose and commitment.
  • Read more

7. Salesforce Just Reaccelerated Growth at $45B ARR — SaaStr

  • Why read: Validation of the agentic software model driving enterprise revenue growth.
  • Summary: Salesforce broke $1 billion in ARR with its Agentforce product line. It is the fastest-scaling launch in the company's history. This proves enterprises will pay a premium for AI agents that execute complex workflows rather than simple conversational interfaces. Agentforce shows how mature software companies can use AI to reignite growth. The B2B recovery is rewarding platforms that integrate autonomous AI into their core products.
  • Read more

8. A rational conversation on where AI is actually going | Benedict Evans — Lenny's Newsletter

  • Why read: A perspective on the AI hype cycle that compares today to the 1997 internet boom.
  • Summary: Benedict Evans argues we are in the "1997" phase of AI: high excitement but uncertainty about future applications and value capture. As AI makes software easier to build, distribution becomes the main competitive moat. AI will likely automate specific tasks rather than entire jobs, creating a boom in professional services as companies figure out implementation. Operators should focus on distribution and understanding how AI reshapes specific workflows.
  • Read more

9. How to Share a Vision You Haven't Finished Building — Regina Gerbeaux

  • Why read: Advice for leaders trying to align teams during strategic uncertainty.
  • Summary: Founders often go silent when their vision isn't clear, causing anxiety and misalignment. Waiting for the perfect answer is a mistake. Leaders need to share their working hypotheses and decision-making frameworks, even if the final destination is blurry. Articulating the current state of the vision and what you are exploring maintains trust and keeps teams focused during strategic pivots.
  • Read more

10. The fallacy of this is that more creates more — Karri Saarinen

  • Why read: An argument against "grindmaxxing" culture in startups.
  • Summary: Believing that more hours equals better output is a fallacy, especially in startups where advantages come from strategy and novel approaches. If hours were the deciding factor, large corporations would always win. Grinding diminishes the perspective needed for creative problem-solving. Long-term success comes from creating conditions for quality work like rest, space, and clear thinking, not daily heroics.
  • Read more

11. Less Debate, More Deployment — Jigar Shah

  • Why read: An argument for prioritizing execution over policy analysis in reindustrialization.
  • Summary: The U.S. needs to shift from diagnosing manufacturing decline to deploying projects at scale. Hollowing out the industrial base was a policy choice; rebuilding it requires fixing specific institutional machinery and procurement systems. China's dominance in clean energy comes from strategic national interest, not climate diplomacy. Operators and policymakers need to prioritize execution and speed to bridge the gap between invention and production.
  • Read more

12. The Magical Money Tree of Management Fees — Dan Gray

  • Why read: A breakdown of how venture capital management fees are distorting startup exit timelines.
  • Summary: As venture capital funds grow, management fees outpace returns from later-stage investments. This incentivizes firms to deploy larger rounds and delay startup exits to justify their balance sheets. Subsidizing unit economics to hit growth metrics pushes profitable outcomes further out. The venture ecosystem's appetite for fees distorts the economics and timelines of building sustainable businesses.
  • Read more

13. SITALWeek #463 Stuff I Thought About Last Week — Brad Slingerlend, NZS Capital

  • Why read: A framework for identifying power-law winners in the AI transition.
  • Summary: The market is dominated by a few trillion-dollar tech companies, showing the power-law dynamics of technological shifts. As AI reshapes the economy, winners will be those who create network effects and vertical integration. Adaptability and creating more value for the ecosystem than is extracted will separate enduring platforms from the rest. Evaluate companies on these traits rather than current revenue.
  • Read more

14. Entrata S-1 | An Actual Software IPO?! — OnlyCFO's Newsletter

  • Why read: What the Entrata IPO filing means for the broader software market.
  • Summary: Entrata's S-1 filing for a property management software IPO signals that the market for traditional, non-AI software companies remains viable. Built efficiently and without massive venture funding, Entrata shows that bootstrapped growth metrics can command public market interest. This IPO could open the door for a backlog of private software companies waiting to go public. Durable revenue and operational efficiency are still valued, despite the AI frenzy.
  • Read more

15. Reflecting on the 200 series — David Hoang

  • Why read: How software craftsmanship and architecture are changing as AI commoditizes code.
  • Summary: As AI commoditizes code generation, human craft becomes the main differentiator for product success. Software architecture is shifting from isolated walled gardens to highly interoperable systems, APIs, and partnerships. Design systems are becoming inference systems, changing how interfaces interact with capabilities. Operators who adapt to these architectural shifts while focusing on product quality and craftsmanship will hold an advantage.
  • Read more

Themes from yesterday

  • The shift in AI from raw model capabilities to optimizing harnesses and compiling agent workflows.
  • Re-evaluating the ROI of frontier models as performance gains shrink and enterprises look for cheaper alternatives.
  • Pushback against grinding culture and AI-fueled distraction, focusing instead on craft, clear vision, and sustainable engineering.
  • The B2B landscape is evolving: high AI platform valuations sit alongside traditional software IPOs grounded in operational efficiency.