AI Builders Digest

Anthropic traces Claude Code's quality dip to three separate changes

Three unrelated changes stacked into what looked like broad degradation: the default reasoning effort dropped from high to medium on March 4, a caching optimization shipped March 26 kept clearing Claude's prior thinking on every turn instead of once, and an April 16 system prompt line capping responses at 100 words cost real coding quality. The API was never affected, and all three were resolved by April 20 in v2.1.116. An ablation showed a 3% eval drop from the verbosity instruction alone, and the thinking-clearing bug drove the cache misses behind reports of usage limits draining faster than expected. Usage limits are being reset for all subscribers, and future system prompt changes now get per-model eval suites, soak periods, and gradual rollouts.

• #products
• #evals

Cheaper DeepSeek and Grok models will pull enterprise demand up, not down

Same-day releases from DeepSeek and Grok both delivered large capability jumps at very low cost, and the effect on enterprise demand runs opposite to what falling prices suggest. Enterprises already have vastly more agent use cases than budget: scanning codebases for security issues, reviewing all their documents, processing information across large parts of their workflows. Every drop in cost or rise in capability unlocks another tier of those. It also raises the value of sitting at the applied layer, since more model choice, especially models tuned for different types or costs of work, means more to route and optimize per task.

• #agents
• #products

Claude sessions now follow you from Chrome to desktop, web, and mobile

Claude in Chrome sessions now carry over to desktop, web, and mobile, with conversations saved and skills and connectors working inside the browser. The side panel runs the same Claude Cowork session as the other apps, so sessions live with your account rather than on any single device. Available on Max and Team today, rolling out to Pro in the coming weeks. Anthropic also flagged that browser agents can be tricked by instructions hidden in a page, and published habits users should adopt on top of its own defenses.

• #products
• #agents

Managed Agents can now run tools inside your own perimeter

Claude Managed Agents can execute tools in a sandbox you control, either on your own infrastructure or through Cloudflare, Daytona, Modal, or Vercel. The agent loop for orchestration, context management, and error recovery stays on Anthropic's infrastructure while code execution, sensitive files, and services stay inside your boundaries, with resource sizing and the runtime image set on your side. MCP tunnels, in research preview, reach private MCP servers through a single outbound connection from a lightweight gateway, turning internal databases and ticketing systems into callable tools with no inbound firewall rules and no public endpoints. Self-hosted sandboxes are in public beta, with Amplitude, Clay, and Rogo named as early builds on the pattern.

• #agents
• #products

Splitting the agent brain from its hands cut p50 time-to-first-token 60%

Managed Agents started with session, harness, and sandbox in one container, which made that container a pet: if it died the session died with it, and debugging meant shelling into a box that also held user data. The fix was to virtualize each piece, so the harness calls the sandbox the way it calls any other tool, and a crashed harness reboots with wake(sessionId), pulls the durable event log, and resumes from the last event. Containers now get provisioned only when a session actually needs one, which dropped p50 time-to-first-token roughly 60% and p95 over 90%. The security payoff is structural: credentials sit in a vault or get wired into the local git remote at init, so tokens are never reachable from the sandbox where Claude's generated code runs.

• #agents

Gemini adds a new wave of integrations, from OpenTable to Ticketmaster

Another batch of Gemini app integrations starts rolling out today: Angi, Fever, GetYourGuide, Granola, iHeartRadio, Localiza, OpenTable, Otter, Pandora, Thumbtack, Ticketmaster, Wix, Zocdoc, and Zoho. The spread across booking, ticketing, notes, media, and local services points at Gemini as the place tasks get finished rather than answered. Google is openly recruiting the next round of partners.

• #products
• #agents

Model reasoning has run ahead of the products built on top of it

The gap between what models can actually do and what gets delivered to users is now the industry's real work, a capability overhang that pushed scaling laws off the list of things worth arguing about this year. Making agents useful takes an ecosystem shaped like the internet, with MCP serving the purpose HTTP serves and NL Web the purpose of HTML, plus agent identities and entitlement systems so an agent can request permission for exactly the systems a task requires. 'I want all of our systems internally to speak a standard protocol to all of the agents that we're writing inside of Microsoft,' which is how you avoid shipping your org chart to everyone else. The next shift is synchronous prompting giving way to asynchronous delegation, where an agent goes off, makes many calls, iterates for a long while, and comes back later with how far it got.

• #agents

The next few years of alpha in AI sit in the application layer

Models will keep getting cheaper, better, and more local, so differentiation shifts to whoever understands a specific user workflow deeply enough, and has the imagination to redesign the experience around it. The pool of people who can build AI products is about to get enormous, which is precisely why being in the top 0.1% of builders matters more rather than less. Separately, a lament that years of Zoom and Meet beat the whiteboarding instinct out of teams, along with the playful half-baked drawing and interrupting that used to shape ideas.

• #products

GBrain v0.45.6.0 adds 17 skills and now runs on Codex and Claude Code

GBrain shipped v0.45.6.0 with 17 new brain skills, hardened through a personal OpenClaw agent running over hundreds of thousands of markdown files, and it now works with Codex and Claude Code. The usage guidance matters as much as the release: run it as a separate agent, not inside your main coding agent. Think of it as a personal version of ChatGPT or Claude that keeps its own git repo for memory and custom skills. The direction of travel, in his words: we are gonna be markdown skill-maxxing from here.

• #agents
• #products

Codex ships Linux support, with a fast mode landing and limits reset

Codex just shipped Linux support, a /fast mode was landing within the hour, and usage limits got a reset for everyone. A 15M milestone passed a few days back, mentioned in passing rather than announced.

• #products

Vercel Sandbox now ships with a default toolchain you can customize

Sandbox comes with a reasonable default set of pre-installed tools now, fully customizable, and npx sandbox@latest sh drops you straight in. The claim worth testing: it feels faster than your local machine. Seedance 2.5 is also live on Vercel's AI Gateway.

• #products

The case that voice becomes the orchestration layer for cloud agents

A new essay argues the way we use computers is about to change permanently, moving from manual work with a keyboard, mouse, and laptop to directing agents in the cloud with our voices. Three claims carry it: voice becomes the orchestration layer, personal computers move to the cloud, and trust ends up the key differentiator among the systems doing the work.

• #agents

SPC's first bet outside the US is India

SPC India was the firm's first move outside the United States, placed on the view that some of the most ambitious projects of the decade will get built there. The framing is structural rather than opportunistic: Indian dynamism is here to stay.

• #funding

The agent surface moved from CLI to apps to services in about a year

A compressed timeline of where agent products have landed: CLI was a year ago, apps maybe six months, and now it's services, web, and cloud sessions. Each surface arrived faster than the one before it.

• #agents

A newly shared paper gets called one of the year's most important

The paper is already one of the most important of the year, with the caveat that its methodology is not clearly explained, so it comes with a set of notes and a further distillation rather than a bare link. Also, a scheduled reminder that Perplexity offered to buy Chrome one year ago today.

• #research

Graph engineering is the new loop engineering is the new prompt engineering

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. Hope this clarifies any confusion. The joke lands because each of those labels briefly felt like a discipline before the next rename arrived.

• #agents