AI Builders
Digest

The Takeaway: For agents to be truly useful they need an open, composable ecosystem that looks like the internet, and MCP is already playing the role of HTTP.

Microsoft CTO Kevin Scott argues that the past year produced a surge in agents and time spent with them, yet model reasoning has outpaced what products deliver—a “capability overhang.” Scaling laws are no longer the main debate; the focus has shifted to the agentic web: better agentic memory, tool use, and protocols that let existing websites and APIs talk to agents.

Agents must take action, call tools, and consult rich information sources. Simple open protocols like MCP are emerging as the HTTP of this new web, while NL Web serves a role analogous to HTML. Inside Microsoft, Scott is pushing every internal system to speak a standard protocol to agents so the company does not ship its org chart.

On security he rejects a false dichotomy between open and closed systems. Permissionless innovation is valuable, and robust identity, entitlement, and user-permission flows can be layered on top of MCP in the open. Regarding the craft of coding, he draws on decades as a programmer and woodworker: tools change, people keep strong opinions, and the right response is curiosity—try the new tools and keep what works.

A year from now he expects the shift from synchronous prompting to asynchronous delegation, where agents iterate for long stretches before handing results back.

“Be curious. Try stuff. And if it works for you, use it. And if it doesn’t, don’t.”

Anthropic Engineering: An update on recent Claude Code quality reports. Over the past month users reported degraded responses in Claude Code, the Agent SDK and Cowork; the API was unaffected. Three separate changes were identified and fixed by April 20 (v2.1.116). Usage limits were reset for all subscribers.

First, on March 4 the default reasoning effort was lowered from high to medium to cut long-tail latency; intelligence suffered and the change was reverted April 7. Second, a March 26 caching optimization meant to clear older thinking only on idle sessions had a bug that cleared thinking on every subsequent turn, producing forgetfulness and repetition; fixed April 10. Third, an April 16 system-prompt instruction limiting text between tool calls to ≤25 words and final responses to ≤100 words hurt coding quality and was rolled back April 20.

Because each change hit different traffic slices, the aggregate looked like broad, inconsistent degradation. Internal evals and dogfooding initially missed the issues. Going forward Anthropic will have more staff use the exact public build, improve internal Code Review, run broader per-model evals on every system-prompt change, and use soak periods plus gradual rollouts for any intelligence trade-off.

Anthropic Engineering: Scaling Managed Agents: Decoupling the brain from the hands. Managed Agents virtualizes the session (append-only event log), harness (the loop that calls Claude and routes tools), and sandbox (execution environment) into interfaces designed to outlast any particular implementation—the classic systems problem of designing for “programs as yet unthought of.”

Coupling everything into one container created a “pet”: container failure lost the session, debugging was nearly impossible, and connecting to customer VPCs required network peering. Decoupling lets the harness treat the sandbox as just another tool via execute(name, input) → string; containers become cattle. The session log lives outside the harness so a new harness can wake(sessionId) and resume. Credentials never enter the sandbox—Git tokens are wired into the remote at init, MCP OAuth tokens live in a vault and are fetched by a proxy—so prompt injection cannot simply read environment secrets.

The session is not Claude’s context window; it is a durable, queryable event stream that supports slicing, rewinding and arbitrary context transformations inside the harness. The architecture cut p50 TTFT by ~60% and p95 by >90% and scales to many brains and many hands. Interfaces stay stable while implementations evolve with the models.

Claude Blog: New in Claude Managed Agents: self-hosted sandboxes and MCP tunnels. Self-hosted sandboxes let a Managed Agent execute tools on infrastructure the customer controls (or via Cloudflare, Daytona, Modal, Vercel). Sensitive files, packages and services stay inside the enterprise perimeter while orchestration and context management remain on Anthropic’s side. MCP tunnels use a lightweight customer-deployed gateway that makes a single outbound connection, allowing agents to reach private-network MCP servers without public endpoints. Self-hosted sandboxes are in public beta; MCP tunnels are in research preview.

Box CEO Aaron Levie notes that the latest DeepSeek and Grok model updates deliver huge capability jumps at extremely low cost—classic Jevons paradox for AI. Lower cost expands demand. Enterprises already have far more agent use-cases (security scans across codebases, document review, workflow information processing) than they can currently budget for. Cheaper or more capable models unlock new use-cases and further demand. The applied-AI layer benefits as well: greater model choice, especially models tuned for different work or cost profiles, increases the value of the routing and optimization layer.

Y Combinator President & CEO Garry Tan shipped GBrain v0.45.6.0 with 17 new brain skills hardened through his personal OpenClaw agent and hundreds of thousands of markdown files. It now works with Codex and Claude Code. He stresses that GBrain should run as a separate agent with its own git-repo memory and custom skills—more like a personal ChatGPT or Claude—rather than inside the main coding agent. The direction is clear: markdown skill-maxxing.

South Park Commons General Partner Aditya Agarwal says SPC’s first major bet outside the US was India. He believes some of the decade’s most ambitious projects will be built there and that Indian dynamism is here to stay.

Vercel CEO Guillermo Rauch highlights npx sandbox@latest sh as feeling faster than a local machine, with a sensible default toolset that is fully customizable. Seedance 2.5 is now available on the Vercel AI Gateway. His broader take: “Endless opportunity everywhere you look.”

Claude announced that Claude in Chrome sessions now carry over to desktop, web and mobile; conversations are saved and skills/connectors work in the browser (Max and Team today, Pro in coming weeks). The side panel runs the same Claude Cowork session across devices. Claude also warned that browser agents can be tricked by instructions hidden in a page and shared recommended defensive habits.

Google VP Josh Woodward announced another wave of Gemini integrations with everyday apps: Angi, Fever, GetYourGuide, Granola, iHeartRadio, OpenTable, Otter, Pandora, Thumbtack, Ticketmaster, Wix, Zocdoc, Zoho and more. Partners interested in integrating can DM him.

OpenAI Codex & ChatGPT’s Thibault Sottiaux noted the project has crossed 15 M and quipped “Also don’t say Linux, we just shipped that.” He also shared enthusiasm for the core-alignment work and what the team is building.

Swyx (smol.ai and others) called a recent paper “already one of the most important papers of this year” and shared distilled methodology notes. He also marked the one-year anniversary of Perplexity’s reported offer to buy Google Chrome.

Peter Yang argues in a new essay that how we use computers is about to change forever—from manual keyboard-and-mouse work on a laptop to directing agents in the cloud with our voices. Key claims: voice becomes the orchestration layer, personal computers move to the cloud, and trust will be the decisive differentiator.

Meta AI Senior Director Madhu Guru contends that the biggest alpha in AI products over the next few years sits in the application layer. Models will keep getting cheaper, better and more local; differentiation will come from deeply understanding specific user workflows and having the imagination to redesign the experience around them. The denominator of people who can build AI products is about to become enormous, so being in the top 0.1% of builders will matter more than ever.

Peter Steinberger (OpenClaw) observed the rapid shift: CLI was a year ago, apps maybe six months ago, and now it is services, web and cloud sessions.

Zara Zhang called a Stanford lecture series “pure gold” and noted how remarkable it is that such high-quality knowledge is freely available on YouTube.

FirstMark VC Matt Turck quipped that graph engineering is the new loop engineering, which is the new harness engineering, which is the new context engineering, which is the new prompt engineering—hoping that clarifies any confusion.

Anthropic philosopher and ethicist Amanda Askell noted the varying bars for “ethical” playthroughs: in Bioshock the bar is literally not murdering children while murdering adults is fine. In her own Skyrim run she largely avoided killing, adopted orphans and built them houses, remembering the plot mainly as “that challenging fantasy philanthropy game.”

Meta’s Madhu Guru misses the playful energy of standing around a whiteboard, sketching half-baked ideas, interrupting each other and jointly shaping things—an instinct that years of Zoom and Google Meet seem to have beaten out of teams.

FPV Ventures partner Nikunj Kothari again called Ethan Ding a “mad mad genius” and pointed to his latest post as a good reminder.