1. [AINews] Anthropic raises $965B Series H, releases Opus 4.8 and Dynamic Workflows/ultracode — AINews

  • Why read: Anthropic is operating at a scale that breaks traditional venture math.
  • Summary: Anthropic confirmed a $65B funding round on a $900B valuation and hit a $47B revenue run-rate. They released Opus 4.8 to fix recent quality issues and set new coding benchmarks. The biggest update is "ultracode," a feature allowing Claude to execute tasks in parallel. It reportedly rewrote the 750,000-line Bun runtime from Zig to Rust in six days. We are moving from chat assistants to actual engineering engines.
  • Read more

2. What we’ve learned building cloud agents — Josh Ma (Cursor)

  • Why read: Local agents are easy because they use your setup; cloud agents are an infrastructure nightmare.
  • Summary: Cursor found that agent quality depends on the environment, not just the model. Building agents in the cloud requires solving hibernation, VM forking, and security at scale. They moved their agent loop to Temporal so tasks can run for days without failing due to infrastructure blips. This makes agent building more of an OS and orchestration problem than a modeling one.
  • Read more

3. Under the River — Burke Libbey (Shopify)

  • Why read: How Shopify got agents to write 12% of their code.
  • Summary: Shopify’s "River" agent now handles one in eight merged pull requests. They succeeded by cleaning up their infrastructure first, moving to a monorepo and using Nix for reproducible builds. This made the code legible for machines. The takeaway: if your codebase is easy for a human to navigate, it is easy for an agent. Centralized agents that see every interaction eventually outperform local ones.
  • Read more

4. Last call for tokenmaxxing — Michael Mignano

  • Why read: The era of unlimited model spend is ending.
  • Summary: The days of using frontier models for every basic task are over. We are entering an "AI FinOps" phase where companies build custom harnesses to trim context and cache results. Expect a shift where grunt work goes to cheap open-source models, while expensive frontier models are reserved for hard logic. You need budget tracking in your agent stack now.
  • Read more

5. Your team needs a unified MCP. Here’s a recipe. — Basis

  • Why read: A blueprint for giving non-technical teams 5x more AI tools.
  • Summary: Basis built an internal gateway called "Satellite" to centralize tool access. By handling auth via Google SSO, they saw average tool usage jump from 3 to 17 per person. This levels the field for non-engineers and lets security teams enforce policies in one place instead of monitoring every laptop.
  • Read more

6. The App Layer is Dead. Long Live the App Layer — Siddharth

  • Why read: Why most AI wrappers will fail.
  • Summary: Simple task boards and notes are being commoditized by models. To survive, software needs to tie directly to a metric the customer cares about, like conversion or throughput. A "wrapper" just shows off a model; a "product" uses that model to change a business outcome. The winner is the system with the best feedback loop, not the best UI.
  • Read more

7. Agent Sandbox Infra on K8S and AWS — Brian Giori

  • Why read: The technical stack needed to let agents run code safely.
  • Summary: Running agent code requires a real sandbox. Amplitude uses short-lived microVMs on Kubernetes where secrets are injected via proxy and networks are locked. Their "SandboxBackend" manages the lifecycle through a custom controller. If you are building an internal agent layer, this is the infra blueprint.
  • Read more

8. How to build your own agent harness??? — Mike Piccolo

  • Why read: Why you should stop using LangChain and build a modular harness.
  • Summary: Teams usually outgrow frameworks like LangChain once they need custom policies or memory. A better approach is building a harness of small, independent workers—routers, budget trackers, and credential vaults—connected by a shared engine. This modularity lets you swap parts without a full rewrite. It also makes production observability easier via OpenTelemetry.
  • Read more

9. The Last Year of Learning AI. — Eira

  • Why read: Prompt engineering courses are mostly procrastination.
  • Summary: In 2026, knowing how a transformer works will not give you an edge. The gap is now between people who ship and people who take courses. Prompt engineering is becoming a commodity; what matters is your judgment and taste. Stop studying and ship five real projects this week.
  • Read more

10. The First Supercompany? — Greg & Taylor

  • Why read: How Anthropic generates $10M in revenue per employee.
  • Summary: Anthropic hit a $30B revenue run-rate with 3,000 people. Google needed 10x the headcount to reach that same milestone. This marks the rise of "Supercompanies" that use AI for internal operations instead of just the product. Their engineers use parallel agents to ship up to 30 PRs a day. Headcount is no longer the way to measure scale; its AI-native operating structure is.
  • Read more

11. Visa invests in Replit to power agentic payments for developers — Ivan Mehta

  • Why read: Agents are getting their own credit cards.
  • Summary: Visa is investing in Replit to build a "Trusted Agent Protocol." The goal is to let AI agents identify themselves and pay for services directly. This closes the loop on economic transactions. Agents will not just write code; they will buy the APIs and compute they need to finish the job.
  • Read more

12. The Age of Async Agents — Cognition's Walden Yan & OpenInspect's Cole Murray — Latent.Space

  • Why read: Moving from chatbots to background engineering teammates.
  • Summary: We are moving from chat assistants to "spec-to-PR" agents that work in the background. These agents need their own VMs and persistent memory to handle multi-day engineering tasks. While anyone can build a simple agent now, the real challenge is the deep reasoning required to get high commit rates.
  • Read more

13. Security in the Age of AI Agents: Office Hours with Jonathan Jaffe — Tomasz Tunguz

  • Why read: Security is becoming an automated engineering task.
  • Summary: With code volume exploding, security cannot be a manual check. At Lemonade, agents already scan threat intel and verify if vulnerabilities are reachable in production. Every agent now needs a unique identity and specific permissions. The "window of exploitability" is shrinking because AI can patch systems faster than humans can find bugs.
  • Read more

14. Centrally Managed Shared Agents — Zara Zhang

  • Why read: Why companies are killing "Shadow AI" by buying team-wide agent seats.
  • Summary: Companies are moving away from individual AI subscriptions and toward centrally managed agent pools. This kills "Shadow AI," where employees pay for their own tools, and ensures data security policies are followed. It also makes it easier for non-technical teams to start using shared skills developed by the rest of the company.
  • Read more

15. How Owner.com's CRO Is Closing $2M Per Rep With AI — SaaStr

  • Why read: How to close $2M in sales per rep using AI agents.
  • Summary: Owner.com is closing $2M in ARR per sales rep by using agents for research and qualification. They did not just automate emails; they built agents that handle the heavy lifting of the sales cycle. This is the new benchmark for GTM efficiency: high revenue with a small, augmented team.
  • Read more

Themes from yesterday

  • The 10x Headcount Rule: Companies like Anthropic are proving that AI-native operations can hit massive revenue with 90% fewer people than previous tech giants.
  • Agents as Teammates: We are past chat windows. The new standard is background agents on dedicated VMs with persistent memory and the ability to execute code.
  • The AI FinOps Era: Token euphoria is ending. Teams are centralizing agent management and building custom stacks to control costs and security.
  • Metric-First Products: Software that merely uses AI is dying. Winners tie model capabilities directly to business outcomes like conversion or throughput.