AI Builders
Digest

The Takeaway: AI automation doesn't eliminate human work—it multiplies it by making yesterday's expertise cheap and widely available, creating more demand for experts to refine and innovate. Dan Shipper, co-founder of Every, draws from his company's experience growing from 4 to 30 people while heavily using AI agents and tools like Cloud Code. He argues that flooding systems with "close but not quite right" AI outputs increases the need for human experts to build better systems and create novel work. "The further away an agent gets from a human, the less valuable it is." Shipper emphasizes that even as models improve exponentially, benchmarks can be easily "unsaturated" with new framings, and humans remain essential for deciding what matters in a rapidly changing world. This philosophy offers hope amid doomer narratives: ride the models, and ambitious work becomes more accessible.

Anthropic Engineering: An update on recent Claude Code quality reports. The team investigated user reports of degraded performance in Claude Code, tracing issues to three changes: default reasoning effort lowered from high to medium (reverted), a caching optimization bug causing forgetfulness (fixed), and a system prompt to reduce verbosity that hurt coding quality (reverted). All issues resolved by April 20. They emphasize taking feedback seriously and are resetting usage limits. Key lesson: even small prompt and configuration changes can have outsized effects. https://www.anthropic.com/engineering/april-23-postmortem

Anthropic Engineering: Scaling Managed Agents: Decoupling the brain from the hands. Managed Agents is a hosted service for long-horizon agents with stable interfaces for session, harness, and sandbox that can evolve independently. By decoupling the "brain" (Claude + harness) from "hands" (sandboxes/tools) and session log, they achieved better reliability, security, and performance (p50 TTFT dropped 60%). Supports many brains and many hands. This architecture prepares for future model improvements without breaking existing systems. https://www.anthropic.com/engineering/managed-agents

Claude Blog: New in Claude Managed Agents: self-hosted sandboxes and MCP tunnels. Users can now run agent execution in their own infrastructure or managed providers (Cloudflare, Daytona, Modal, Vercel) while keeping orchestration on Anthropic's side. MCP tunnels allow secure connection to private internal services without public exposure. Enhances enterprise security and control for sensitive workloads. https://claude.com/blog/claude-managed-agents-updates

Claude Blog: New connectors in Claude for everyday life. Expanded directory now includes AllTrails, Instacart, Audible, Tripadvisor, TurboTax and more lifestyle apps. Connectors dynamically suggest relevant tools in conversations. All ad-free with user control. https://claude.com/blog/connectors-for-everyday-life