1. Coinbase’s bet on agent-first development — linear.app

  • Why read: Shows how Coinbase redesigned its workflows around AI agents instead of bolting them onto human processes.
  • Summary: Coinbase engineers paused coding for two weeks to redesign their development cycles around agents. They replaced scattered Slack conversations with Linear to give agents a structured source of truth. Now, custom agents like Forge handle bugs from report to pull request, while others answer status queries. This shifts the engineering role from writing code to directing agent swarms in a continuous 24/7 loop, sharply reducing coordination friction.
  • Read more

2. The Principal-Agent vs. Agents Problem — X (formerly Twitter)

  • Why read: Explains why companies struggle to capture the value of AI productivity gains.
  • Summary: If AI automates tasks but revenue stays flat, employees capture the surplus as leisure time while the enterprise gains nothing. Employees have a rational incentive to hide efficiency gains to avoid more work or layoffs. To capture economic value, companies must redesign workflows for higher output, sell completed outcomes instead of employee time, and share the upside with workers to encourage transparency.
  • Read more

3. Subagents on Subagents: How Many Layers Deep Is Too Many? — X (formerly Twitter)

  • Why read: Offers a mental model for managing the risks of recursive agent delegation.
  • Summary: When agents spawn subagents, workflows become complex graphs. The main engineering challenge is managing the blast radius of errors: mistakes from upstream agents can establish false premises that downstream agents amplify. Treating agent orchestration as dependency management allows teams to enforce strict verification or human review at high-impact nodes, controlling how uncertainty spreads through the system.
  • Read more

4. Patterns and problems in emerging multiagent systems — anthropic.com

  • Why read: Details coordination failures when autonomous agents interact in shared environments.
  • Summary: Anthropic found that agents can parallelize independent tasks like vulnerability scanning but struggle with interdependent projects like software development. One common failure mode is conformity: agents sharing similar contexts and underlying models tend to make identical errors, causing systemic collapses. While newer models improve coordination by minimizing code sharing, true collaboration remains difficult. Builders must solve these behavioral homogenization issues to make multiagent systems work.
  • Read more

5. Enterprise Signals | OpenAI — openai.com

  • Why read: Data showing the enterprise shift from chat to agents and the growing gap between top AI users and typical firms.
  • Summary: OpenAI data shows over 60% of enterprise output tokens now come from agentic workflows instead of chat. The top 10% of companies generate eight times more output per user than average firms by giving agents tools, context, and persistence to execute multi-step tasks across legal, sales, and marketing. Early-career employees are driving this shift. Average firms need to move past basic chatbots and embed AI directly into operations to catch up.
  • Read more

6. How I delegate the work, not my judgment — Depot

  • Why read: A guide to keeping system ownership and architectural intuition when using coding agents.
  • Summary: Delegating code to agents can erode a developer's understanding of their own systems if they only review final outputs. Engineers must stay active in the planning phase to set architectural tradeoffs and scope. Reviewing AI code requires focusing heavily on long-term maintainability, abstraction, and naming rather than simply checking if tests pass. Limiting parallel agent tasks ensures the developer retains control over the codebase's evolution.
  • Read more

7. Agent Plugins are the future of Agent Skills — X (formerly Twitter)

  • Why read: Outlines a new open standard for distributing AI agent capabilities.
  • Summary: Sharing agent skills usually requires complex setups linking instructions to separate servers. The new Agent Plugins standard bundles instructions, tools, and MCP servers into a single portable folder. This vendor-neutral format lets tools and instructions travel together across compatible clients. Components fail independently, so a broken server won't crash the agent. Standardizing this distribution lowers the friction for teams sharing workflows.
  • Read more

8. Cracks in the AI Thesis — Ramp

  • Why read: Spend data showing businesses hitting a price ceiling on frontier AI models and shifting to open-source.
  • Summary: Business spending on expensive frontier models like Anthropic's Fable 5 is lagging. Fable 5 accounts for only 6% of tokens bought from Anthropic, showing that marginal performance gains aren't worth the premium for most enterprises. Meanwhile, advanced AI spenders are increasingly using platforms that host cheaper open-source models. As open-source closes the capability gap, major AI labs are losing pricing power. Infrastructure companies now have to prove clear value to justify their costs.
  • Read more

9. The 24/7 simulator that finds our bugs before customers do — X (formerly Twitter)

  • Why read: Shows how Phoebe uses AI agents to simulate user behavior and catch edge cases in production.
  • Summary: Phoebe built a 24/7 simulation environment to test its AI agents. The simulator uses 20 AI personas with distinct traits to interact with the product like real users, generating calls, texts, and scheduling conflicts. Running on production infrastructure, it finds logic errors and regressions over days and weeks that unit tests miss. It also acts as a safe staging ground for feature rollouts and provides a populated demo environment for sales.
  • Read more

10. How a GPU Actually Works — X (formerly Twitter)

  • Why read: Explains why memory bandwidth, rather than arithmetic power, limits LLM inference speed.
  • Summary: GPUs have massive arithmetic power but are bottlenecked by memory retrieval speeds. Generating a single token requires reading the entire model weight into memory for very few calculations, making text generation highly inefficient. Engineers improve performance by increasing the "work per byte" ratio using batching, which processes multiple requests with the same loaded weights. Techniques like quantization and operator fusion also help by reducing the total bytes fetched.
  • Read more

11. There’s so much to unpack here, but it’s a huge... — X (formerly Twitter)

  • Why read: A warning on the risks of building a business dependent on third-party integrations.
  • Summary: Software integrations are revocable treaties, not permanent features. When a platform controlling a system of record sees a partner capturing a valuable workflow, it often revokes access to take that value for itself. Platform-dependent revenue is fragile and investors discount it heavily. Startups need to build data portability and operational continuity before they get cut off. Customers suffer most during vendor divorces, which is why procurement should prioritize continuity over connectivity.
  • Read more

12. AI and organic search are doing different jobs: What Shopify's data shows — Shopify Plus

  • Why read: Shopify data showing how AI search changes e-commerce discovery.
  • Summary: Shopify's Q2 data shows AI search drives 197% year-over-year growth in referred sessions, serving a different need than organic search. Shoppers use AI for research-heavy, spec-driven purchases where they weigh constraints, and this traffic converts at double the rate of organic search. Organic search still dominates broad, taste-led browsing. Optimizing for AI search mostly requires structuring product data cleanly. Merchants should enrich their catalogs with detailed specs and conversational attributes to capture this high-intent traffic.
  • Read more

13. Agent Adoption — Chris’s Substack

  • Why read: Examines why AI agents lack consumer adoption and how their interfaces need to evolve.
  • Summary: Mainstream adoption of AI agents has stalled because they require too much setup, break easily, and rely on text chat. Chatbots force a high cognitive load on users to generate prompts, making them poor tools for daily tasks. Agents need to evolve into multimodal, dynamic interfaces that visually guide users. Developers must focus on building reliable, background agents that solve immediate problems with near-zero friction.
  • Read more

14. Ok friends, sharing the recipe to build basically any agent... — X (formerly Twitter)

  • Why read: A breakdown of the architecture powering proactive AI agents.
  • Summary: Building an effective agent requires a standard stack: memory of user context, clear operational skills, and tools to access external systems. The core is a runtime loop that plans, acts, and delegates, triggered by cron jobs or system events instead of user prompts. The new competitive frontier is proactive background agents that execute tasks independently. Success depends on balancing high-quality execution with low inference costs to preserve gross margins.
  • Read more

15. Hiring Agents Is the Easy Part — X (formerly Twitter)

  • Why read: Explains the operational challenges of enterprise AI through the lens of employee management.
  • Summary: As companies shift from AI augmentation to fully automated workloads, they face challenges similar to managing human employees. Screening models and connecting them to company tools is getting easier, but evaluating their ongoing performance is difficult. The real bottleneck to scaling AI is verifying the quality of nuanced work and providing feedback that helps agents learn. Companies also need to resolve questions around liability, data ownership, and internal permissions.
  • Read more

Themes from yesterday

  • Agents in the Enterprise: Companies are moving past chat interfaces and adopting proactive background agents to run workflows autonomously.
  • Chat UI Limits: Conversational interfaces create too much friction for mainstream adoption. The next phase of agents requires dynamic, multimodal applications.
  • AI Economics: Enterprises are hitting a price ceiling on frontier models and shifting toward cheaper open-source alternatives to balance cost and performance.
  • The Engineering Pivot: As AI writes more code, engineers are shifting focus to architectural judgment, strict verification, and managing multi-agent dependencies.