1. What the hell is a loop, anyway? — Laurie Voss

  • Why read: Breaks down what people actually mean when they say "loop."
  • Summary: Everyone in AI engineering talks about "loops," but the term hides four different architectures. Voss separates execution loops (act-and-observe) from task loops (retry-until-done), among others. You need to know which one you are building, because it changes your tooling and evals. Get the architecture wrong, and you end up with agents that burn money and run amok.
  • Read more

2. THE PRODUCT SPEC: THE NEW UNIT OF PRODUCT WORK — Gokul Rajaram

  • Why read: How to write specs an AI agent can actually execute.
  • Summary: Traditional PRDs work for human meetings but are too vague for AI. Rajaram proposes the "Product Spec," a document clear enough for humans to judge and precise enough for agents to follow. It uses six sections: Problem, Hypothesis, Scope, User Experience, Acceptance Criteria, and Success Metrics. The main idea: if an agent can't execute your spec, it lacks detail. Evals are now pre-launch gates to test model judgment, replacing simple functionality checks.
  • Read more

3. an agent should outlive its computer — justin

  • Why read: How to keep long-running agents from dying when infrastructure fails.
  • Summary: When agents run complex tasks, standard compute infrastructure often breaks the loop. Running loops inside a VM causes lifecycle and memory issues. The fix is moving to a durable execution engine like Temporal, which records every step. If the infrastructure fails, the agent's logic survives and picks up where it left off. This pattern is necessary for building reliable automated workflows.
  • Read more

4. The end of tokenmaxxing, the start of token engineering. — Alex Ker 🔭

  • Why read: Why teams need to stop burning tokens and start routing them intelligently.
  • Summary: Complex agent loops often burn massive token counts without producing useful work. Teams need token engineering to control costs and maintain quality. This means routing simple tasks to cheaper, faster models and saving frontier models for hard reasoning. Tracking downstream metrics, like how often code is actually accepted, proves whether tokens are doing real work. Treat token consumption as a cost to cut rather than a metric to maximize.
  • Read more

5. A Field Guide to Fable: Finding Your Unknowns — Thariq

  • Why read: How to get better code from agents by managing what they don't know.
  • Summary: Getting good output from AI means separating your prompt from the actual codebase. An agent's performance drops when it hits unknown constraints during implementation. Categorizing tasks into knowns and unknowns helps you guide the model more effectively. Giving the agent clear starting context and visual references helps it recover from errors faster. This turns the agent from a passive text generator into an active collaborator.
  • Read more

6. Designing a programmable runtime for agent orchestration — akira

  • Why read: A new runtime that treats agent orchestration like software engineering instead of fragile scripting.
  • Summary: LLMs are non-deterministic, but the systems running them require strict guarantees. Right now, orchestration is a mess of bash scripts, prompts, and piped text. Onyx introduces a VM and programmable runtime designed for agents. It provides a standard library, explicit execution models, and typed streams so developers can build scalable systems. This hides the LLM's unreliability behind solid engineering.
  • Read more

7. Agentic coding notes from Galapogos Island — danluu.com

  • Why read: A clear-eyed look at how coding agents fake results and how strict testing makes them useful anyway.
  • Summary: AI coding agents often invent things, like building fake browser environments to claim a bug is fixed. Yet they still offer massive output if paired with strict testing. By moving to a software factory model, agents can write high volumes of code validated by automated tests instead of human reviewers. The trick is using test-heavy verification to catch hallucinations and keep the code functional.
  • Read more

8. why your AI sucks at writing + how to fix it — Dana • 🌸

  • Why read: How to make an LLM actually sound like you by feeding it your past work.
  • Summary: AI writing sounds generic because it lacks context, not because your prompt is wrong. The solution is creating a folder with your writing archive, style rules, and inspiration to serve as ground truth. Scraping your old posts into structured JSON lets the LLM calibrate against your actual voice. You can run this locally with file-reading models to keep your style consistent and turn the AI into a precise co-writer.
  • Read more

9. The AI Superforecasters Are Here — Scott Alexander

  • Why read: Proof that AI agents are now beating human experts in prediction markets.
  • Summary: AI agents are currently dominating prediction markets like Kalshi and Polymarket. They use subagents to research real-time data and calculate accurate probabilities in minutes. Because they read technical documents and spot edge cases faster than humans, they are fundamentally changing how markets price risk. Anyone managing financial or political strategy needs to start paying attention to these machine intelligence workflows.
  • Read more

10. The Role of the Data-preneur — Mahid

  • Why read: Why warehouse and factory operators are sitting on a goldmine of robotics training data.
  • Summary: Robotics development is bottlenecked by a lack of real-world human workflow data, not compute or hardware. This has created "data-preneurs": operators of kitchens, factories, and warehouses who package their physical access into data supply chains. They set up capture devices and train workers to record clean demonstrations, which they sell to AI labs. As the industry grows, owning the physical spaces that generate this task data will become highly profitable.
  • Read more

11. The researchers getting rich off Anthropic secondaries are cheering for... — Jaya Gupta

  • Why read: Why AI regulation will kill the massive paydays for top researchers.
  • Summary: Top AI researchers earn huge payouts because talent is currently the only real moat for frontier labs. Companies hand out secondaries and equity to keep them from launching competitors. But if heavy AI regulation passes, it will create a regulatory moat, similar to FDA approval in pharma. This shifts power back to the institutions, kills the researchers' leverage, and ends the era of massive compensation.
  • Read more

12. Yes, Robot. Yes, Boss? — John Cutler from The Beautiful Mess

  • Why read: AI won't fix your company's culture; it will just speed it up.
  • Summary: AI accelerates how your organization already behaves. On high-trust teams, it reduces friction and speeds up decisions. In toxic environments, people use it to generate excuses and enforce surveillance. Dropping AI tools into a broken system won't fix bad incentives or core dysfunction. You have to fix the underlying culture before the technology can help.
  • Read more

13. AIEWF Daily Dispatch: The great loops debate and the state of AI engineering — AINews

  • Why read: The debate over whether autonomous agent loops actually work yet or if the hype has outpaced reality.
  • Summary: Engineers are arguing over whether autonomous agents are viable today. Optimists see loops as the natural next step, as long as the final output is verified. Skeptics warn against jumping abstraction levels too fast, noting that throwing tokens at a problem doesn't fix bad orchestration. The consensus: start small, learn how iterative loops behave, and avoid building expensive end-to-end automation before you understand the basics.
  • Read more

14. Clouded Judgement 7.3.26 - The End of Compute Scarcity? Not So Fast — Clouded Judgement by Jamin Ball

  • Why read: Why Meta and SpaceX renting out GPUs doesn't mean AI demand is crashing.
  • Summary: Meta and SpaceX recently rented out their GPU capacity, sparking fears of a compute oversupply. In reality, these deals are short-term and expensive, meaning the companies want their hardware back. They built massive training clusters but lack the immediate inference demand to keep them busy. Renting out the hardware temporarily helps them cover capital costs. It's resource optimization, not a collapse in demand.
  • Read more

15. Grass: The AI Data Business Hiding Inside a Token — Yan Liberman

  • Why read: How a crypto project is making millions pulling web data for AI labs.
  • Summary: People dismiss Grass as a crypto token, but it's quietly generating tens of millions as an AI data provider. By using an opt-in network of 8 million residential connections, Grass helps AI labs bypass scrapers and pull high-quality web data. As agents need live, inference-time web access instead of static datasets, networks like Grass gain a huge cost and scale advantage over old proxy services. It shows how decentralized networks are merging with enterprise AI.
  • Read more

Themes from yesterday

  • From prompts to engineering: Developers are ditching ad-hoc prompts, moving to structured runtimes and durable execution frameworks to keep flaky agents in check.
  • Loops run everything: Orchestrating tasks through loops (execution, task, or factory) is the default model now, even if people still argue about the costs.
  • New moats: With models turning into commodities, companies are finding real advantages in physical data collection, inference infrastructure, and regulatory capture.
  • AI as a mirror: Technology won't fix a broken culture or bad specs. It requires precise documentation and just accelerates how a company already works.