1. The Agentic Decision Framework: Build, Buy, or Platform — Cannonball GTM

  • Why read: A blunt breakdown of why you are overpaying for software and when to build AI tools in-house.
  • Summary: Software development costs have tanked, yet buyers still pay huge premiums. Take Gong: customers pay upward of $140k a year for transcription, storage, and some basic coaching logic. You can now build the same thing in-house using cheap transcription APIs and models like Claude for a fraction of the cost. Stop signing multi-year contracts with opaque pricing. Audit your tech stack and rethink what you buy versus what you build.
  • Read more

2. Harness engineering for coding agent users — Birgitta Böckeler

  • Why read: How to get reliable code out of agents by building the right constraints and feedback loops.
  • Summary: An agent is a model plus its harness. To get better code and spend less time reviewing it, you need to build an outer harness with two parts: guides (rules and docs) to steer the agent before it writes, and sensors (linters and tests) to check its work after. Use one without the other, and the agent either repeats mistakes or hallucinates. Build both to stop wasting tokens and keep the agent grounded in your codebase.
  • Read more

3. How to keep AI spend flat while token usage grows... — Brian Armstrong

  • Why read: Tactical ways to control API costs as token usage spikes.
  • Summary: Capping usage frustrates developers and slows them down. Control your AI spend through defaults, routing, and caching instead. Default to cheap open-weight models for basic tasks. Use an LLM gateway to route only the hardest planning tasks to expensive frontier models. Cache heavily and keep context windows tight so you stop paying for redundant tokens. Do this right, and usage can scale while costs stay flat.
  • Read more

4. How I AI: Using Codex + Claude Code to Prototype Agents — claire vo 🖤

  • Why read: A smart way to prototype new multi-agent workflows using the coding tools you already have.
  • Summary: Testing complex agent workflows is tedious. Speed it up by using tools like Codex or Claude Code. Have the agent use its own authenticated session to run calls against the product you are building. You can test different agent structures right in the CLI. Once you find a strategy that works, tell the coding agent to draft the technical spec. Then, run a blind test in a fresh thread against a test dataset to validate it.
  • Read more

5. Towards org-level agent harnesses — Lance Martin

  • Why read: What happens when agents move from personal assistants to shared teammates in Slack.
  • Summary: AI is shifting from single-player tools to multi-player agents that live in shared workspaces. This gives new hires instant access to company context. As models gain longer context and better security, we are moving from synchronous commands to asynchronous goals. You can now hand off long-running tasks, like code migrations, to shared agents. Instead of babysitting a script, you collaborate in a channel while the agent reports progress on its own.
  • Read more

6. Clouded Judgement 6.26.26 - Time to Power — Clouded Judgement by Jamin Ball

  • Why read: Why getting power to the data center is the real bottleneck for AI compute.
  • Summary: The biggest hurdle for new data centers is time to power. Well-funded cloud providers can get chips, land, and capital, but they need power secured before they can get debt financing. The problem isn't generating power, but transmitting it. Right now, over 2,300 GW of capacity is stuck in interconnection queues. That is two whole grids' worth of power waiting on bureaucracy and infrastructure. Scaling AI compute means fixing this transmission backlog.
  • Read more

7. What If There Is No Moat Yet? — Tomasz Tunguz

  • Why read: Why AI app founders shouldn't stress over technical moats.
  • Summary: Investors want a technical moat, but for app-layer AI companies, technical leads disappear fast. Moats come in two flavors: leading (new architectures, private data) and lagging (scale, brand, embedded workflows). Infrastructure companies need leading moats. App companies win by executing fast and building lagging moats. Focus on distribution and capturing customer workflows. It is fine to have no moat today if you have a clear plan to build one through execution.
  • Read more

8. AI’s winner-take-all era is over — ashu garg

  • Why read: Why OpenAI won't monopolize the market and intelligence is getting cheaper.
  • Summary: The idea of a single winner in AI is dead. Fierce competition is driving the cost of intelligence to zero, and technical leads disappear fast. OpenAI, Anthropic, Google, and xAI are fighting it out, while open-weight models like DeepSeek offer near-frontier performance for cheap. Add in governments pushing for sovereign models, and the market is fragmenting. The base model is no longer a moat; the real value is in applications, workflows, and infrastructure.
  • Read more

9. The Structural Case for Open-Source AI — Kevin Gee

  • Why read: Why open-source AI is inevitable and caps the pricing power of closed labs.
  • Summary: Open-source models have caught up to closed frontier models in quality, but they remain 70 to 500 times cheaper to self-host than paying API fees. This massive cost gap makes self-hosting the only logical choice at scale. Open models structurally cap the pricing power of closed labs and will compress their margins. Open-source AI will keep spreading, regardless of what regulators try to do.
  • Read more

10. America Invents, China Applies — The Generalist

  • Why read: How China is beating the US at putting AI into production everywhere.
  • Summary: The West is obsessed with building the smartest models, while China is forcing AI into every corner of its economy. Beijing's new 17-point plan pushes AI and consumption into shops, clinics, and schools. It is the old story: America invents, China applies. By subsidizing and mandating AI adoption, China is chasing immediate productivity gains. They do not care who has the best algorithm if they can get a structural efficiency advantage today.
  • Read more

11. Wave the Magic Wand First — Jordan Crawford from On the Edge

  • Why read: Why you should stop tweaking old workflows and build backward from impossible AI outcomes.
  • Summary: Most companies get AI backward, using it to shave seconds off existing tasks. Because AI is brilliant at some things and terrible at others, this incremental approach fails. Instead, picture the impossible outcome you want, then work backward to build it with today's best models. Target entirely new workflows that only top-tier AI can handle. Stop building faster busywork and start creating new value.
  • Read more

12. Human in the /loop — eric zakariasson

  • Why read: How to run multiple AI agents in the background without constantly checking on them.
  • Summary: AI agents are only useful if they run loops in the background while you do other work. First, define a machine-measurable target, like a passing test suite or a specific eval score. Tell the agent to propose changes, measure against the target, and keep or revert the work. Set up a Slack webhook so it only pings you when it finishes, gets stuck, or needs a decision. Run these loops in the cloud, fan out the tasks, and only step in to review the final output.
  • Read more

13. Live draft model training for speculative decoding — Rachel Rapp

  • Why read: How training draft models on the fly makes LLM inference significantly faster.
  • Summary: Speculative decoding uses draft models to speed up LLM inference, but keeping them synced with base models is slow. A new pipeline fixes this by training the draft model in real-time using hidden states extracted from live inference. It runs inside the inference stack and offloads network traffic to the background to avoid latency. By adapting to live traffic without saving offline data, this approach boosted accept rates by 20 percent on average, and over 100 percent for specific patterns. The bottom line is much faster inference at scale.
  • Read more

14. What Should Be Done — Dean W. Ball

  • Why read: Why the US government's messy, unofficial regulation of AI is stalling innovation.
  • Summary: Recent executive actions have created a shadow licensing regime, blocking the release of models like Anthropic's Fable and OpenAI's GPT-5.6. The government has no public safety standards for approval and lacks the technical expertise to evaluate models. Most officials with actual AI experience have been pushed out. Catastrophic AI risks are real, but this opaque, uncoordinated approach just keeps models locked up indefinitely. It kills innovation without actually improving safety.
  • Read more

15. Strange Knowledgeability — Contraptions

  • Why read: A sharp look at how LLMs change human curiosity, just like the first encyclopedia did.
  • Summary: Diderot’s 18th-century encyclopedia gave people a map of reality without requiring them to discover the knowledge themselves. This made many people post-curious, content that all facts were stored in a book. LLMs do the same thing today, acting as an always-on oracle. But offloading facts to an AI does not mean you need to know less. It raises the baseline: you have to know more just to ask the right questions. We are changing how we know things as we co-evolve with our tools.
  • Read more

Themes from yesterday

  • The commoditization of intelligence: Intense competition and open-source models are driving inference costs to zero and killing the moats of frontier labs.
  • Agentic workflow maturation: We are moving from single-player chat to multi-player, org-level agents that run in the background with strict constraints.
  • Physical and regulatory constraints: Grid transmission limits and opaque government licensing are now the main bottlenecks to scaling AI.