1. The AI Agents Stack (2026 Edition) — Paolo Perrone

  • Why read: A map of the six infrastructure layers needed to move AI agents from prototype to production.
  • Summary: The agent stack has grown past simple RAG into a six-layer architecture. The big additions are standardized tool connectivity (MCP), memory as a primitive, and guardrails for real-time behavior. When picking tools, evaluate state management needs, lock-in, and the friction of moving to production. The core unit is still the think-act-observe cycle. Only add complexity when a specific layer breaks. Read this if your team is struggling to scale agents.
  • Read more

2. Coding Is No Longer the Constraint: Scaling Developer Experience to Teams and Agents at Spotify — Spotify Engineering

  • Why read: How Spotify integrated AI coding agents into their internal platforms to accelerate engineering velocity.
  • Summary: Over 99% of Spotify's engineers now use AI coding tools weekly, driving a 76% jump in PR frequency. This was built on Fleetshift, their internal platform for fleet-wide code maintenance. They introduced "Honk," a background coding agent, shifting from deterministic scripts to models for broad code modifications. Getting the most out of AI means pairing capable models with automated internal platforms. Build this infrastructure so agents can handle legacy migrations and maintenance.
  • Read more

3. How I Actually Code (and Review) With AI in 2026 — Christina Lin

  • Why read: How to adjust your coding habits and architecture to work around the limitations of AI agents.
  • Summary: AI agents get confused by "context rot" when overloaded with information. Hidden coupling and implicit initializations wreck their performance. To code well with AI, design for model locality instead of human readability. Enforce strict boundaries: pass dependencies, configs, and environment values explicitly instead of accessing them globally. Precise type signatures act as documentation the AI can trust without reading full implementation files. This keeps the required context window small, cuts down on hallucinations, and improves output reliability.
  • Read more

4. Loop Engineering. — Addy Osmani

  • Why read: A breakdown of loop engineering: building systems that continuously prompt and guide AI agents.
  • Summary: Manual, turn-by-turn prompting is giving way to systems where you set a recursive goal and let the AI iterate. A functional loop needs five things: scheduled automations, isolated worktrees, codified project skills, tool connectors, and sub-agents for specific tasks. These loops rely on external state memory, like a markdown file or ticketing board, because models forget context between runs. Tools like Claude Code have this built in, but you have to watch token costs and quality control. Loop engineering is becoming a baseline skill for deploying agents on complex tasks.
  • Read more

5. Most Developers Do Not Need Agent Loops Yet — AlphaSignal AI

  • Why read: A reality check on when agent loops are actually worth their high token costs.
  • Summary: Automated prompting systems aren't always necessary or cheap. A loop only pays off if the task is highly repetitive, verification is fully automated, your token budget can handle waste, and the agent has senior-level tool access. Without all four, running the loop costs more time and money than it saves. Top models can handle long tasks now, but the orchestration logic is expensive. Check your workflows against these constraints before dropping traditional prompting.
  • Read more

6. Claude Code Dynamic Workflows: What’s Actually Changed — Samuel McDonnell

  • Why read: An analysis of Anthropic's Dynamic Workflows and what they mean for production AI systems.
  • Summary: Claude Code's Dynamic Workflows generate JavaScript orchestration scripts to run up to 16 subagents in parallel, moving away from single-context sequential models. This stops context overflow on large tasks like codebase migrations. The main upgrade is the verification layer: it spins up clean-context reviewers to challenge agent outputs, stopping the LLM from rationalizing its own mistakes. But parallelization multiplies token use, since each subagent needs its own context window. Treat this as an orchestrator of strong checks, and weigh the speed of parallel execution against the high inference costs.
  • Read more

7. Your Agent Harness Should Repair Itself — Akshay 🚀

  • Why read: How AI observability is shifting from passive tracing to automated harness repair.
  • Summary: Current AI observability tools just tell you what failed. They don't fix it or stop it from happening again. When an agent breaks, developers have to manually read traces, write patches, and hope they didn't miss edge cases. The fix is building agent harnesses that repair themselves. Tools like Opik link tracing to an automated coding agent to diagnose failures, propose fixes, verify them, and lock them in as regression tests. Automating this feedback loop is the only way to handle the complexity of production prompts, tools, and updates.
  • Read more

8. ok so i've spent years thinking about what it takes... — claire vo 🖤

  • Why read: A checklist for engineering leaders integrating AI into their team's culture and operating model.
  • Summary: AI adoption takes more than buying licenses. It requires technical, operational, and cultural changes. Leaders need to invest in developer experience and platform teams to tighten the feedback loops agents need. Operationally, sort workflows into explicit buckets: automate, augment, ignore, or kill. Stop delegating tasks and start giving agents defined jobs. Culturally, executives must build and demo AI projects themselves to build fluency and address team fears of replacement. Formalize these changes and cut process bottlenecks to actually multiply productivity.
  • Read more

9. The Substitution Wave in AI — Tomasz Tunguz

  • Why read: How rising frontier model costs and capable open-source options are driving AI buyers to substitute models.
  • Summary: Foundation labs moving into applications, higher prices for frontier models, and capable open-source options are changing AI cost structures. Tech companies are routing prompts to cheaper, open-weight models to hold costs flat while scaling usage. Instead of saving the money, they reinvest it to buy more volume. You have to choose between the high costs of closed models or the efficient unit economics of specialized open ones. Mastering model routing gives you an edge in margins and capabilities.
  • Read more

10. Model Routing Will Control the Future of Economic Value — Tomás Hernando Kofman

  • Why read: Why dynamic model routing is the key piece of infrastructure for managing enterprise AI budgets.
  • Summary: Enterprise AI spending is climbing as developers and agents burn tokens on long-horizon tasks. A token is a unit of compute, not intelligence. Using a frontier model for everything is too expensive. The fix is dynamic model routing: sending each prompt to the cheapest model that can handle it. As open-source models catch up, routing traffic away from frontier models is required to keep costs in check. Mastering this layer is the only way to scale AI deployments sustainably.
  • Read more

11. Someone is Gonna Win a Nobel Prize for Figuring out How Much To Spend on Matrix Multiplication — sam lessin 🏴‍☠️

  • Why read: A look at "Algorithmic Macroeconomics" and the math behind allocating compute.
  • Summary: The industry is focused on basic model routing, but the real problem is calculating exactly how much compute to spend on a prompt. This means factoring in the cost of being wrong, the cost to validate versus generate, and the diminishing returns of higher accuracy. As intelligence becomes a commodity, deciding how much compute a problem deserves requires economic modeling, not simple heuristics. Engineering is shifting toward the economics of compute at scale. Solving this allocation problem is the key to building efficient autonomous systems.
  • Read more

12. Welcome to the Token Casino — Veeral Patel

  • Why read: How gamified UI mechanics and economic incentives drive up token spend in AI coding tools.
  • Summary: AI labs make money when you burn tokens. This conflicts with companies trying to ship efficiently. Tools are adopting casino mechanics: tokens act like chips, status bars simulate flow, and leaderboards gamify consumption. This pushes developers to endlessly retry prompts instead of just writing the code. As execution gets cheaper, the real bottleneck is knowing when to stop prompting and step in manually. Watch out for metered tools designed to maximize idle spend.
  • Read more

13. We Should Take Text Optimization More Seriously — Yoonho Lee

  • Why read: The argument for optimizing prompts, memories, and harnesses instead of relying solely on model fine-tuning.
  • Summary: Machine learning has a bias toward model weights, ignoring the value of optimizing the text layer around the model. But optimizing text works like a gradient update and is much more sample-efficient. If you treat prompts, memory, and retrieval as updateable state, systems can test hypotheses against new evidence before changing. This enables models to learn from single experiences without forgetting past knowledge. Tune this external text layer first before spending money on fine-tuning.
  • Read more

14. I Scored 15,489 Account Executives to Find 135 — for $10. Here's How. — Jordan Crawford

  • Why read: How to use AI to evaluate candidates based on their actual sales environment rather than resume prestige.
  • Summary: Hiring usually fails by ranking candidates on past employer prestige instead of the actual sales environments they worked in. An enterprise rep with nine-month deal cycles is a bad fit for a fast SMB motion, but their resume usually wins. If you define the role across dimensions like deal size, buyer type, and cycle length, you can use AI to score thousands of candidates against those specifics. This programmatic approach strips out the beauty contest bias of normal sourcing and finds overlooked operators. It improves hiring accuracy for a fraction of the cost.
  • Read more

15. The sample efficiency black hole — Dwarkesh Patel

  • Why read: Why data, not architecture, is the actual moat driving AI progress.
  • Summary: AI models haven't gotten more sample-efficient; they just need massive volumes of domain-specific data to get fluent. Progress relies on synthetic data from reinforcement learning, which requires huge amounts of human expert trajectories across specific skills. Producing these expert labels is becoming a massive industry. Because this data can often be pulled from public APIs, open-source models are catching up to frontier models. Data is the ultimate driver here. Understand this dynamic when evaluating if your proprietary data is an actual moat.
  • Read more

Themes from yesterday

  • From prompting to orchestrating: Moving from turn-by-turn prompts to automated loops and self-repairing harnesses.
  • The economics of inference: High token costs, gamified UIs, and the need for model routing to protect margins.
  • Adjusting developer experience: Restructuring internal platforms and architectures to fit AI capabilities and context limits.
  • State over weights: Realizing that optimizing the text layer—memory, state, harnesses—is more efficient than constant fine-tuning.