AI Builders
Digest

Anthropic Engineering: An update on recent Claude Code quality reports

Over the past month Anthropic investigated reports of degraded Claude responses and traced them to three separate changes affecting Claude Code, the Claude Agent SDK, and Claude Cowork. The API was unaffected. All issues were resolved as of April 20 (v2.1.116).

On March 4 the default reasoning effort was lowered from high to medium to cut latency; users preferred higher intelligence and it was reverted April 7. On March 26 a caching optimization intended to clear older thinking after one hour of idle time instead cleared it on every subsequent turn, producing forgetfulness and repetition; fixed April 10. On April 16 a system-prompt instruction to reduce verbosity combined with other changes hurt coding quality and was reverted April 20.

Because the changes hit different traffic slices on different schedules the aggregate effect looked like broad, inconsistent degradation. Anthropic has reset usage limits for all subscribers and is tightening internal dogfooding, code-review tooling, and evals for any change that could trade off intelligence.

“We never intentionally degrade our models.”

Anthropic Engineering: Scaling Managed Agents: Decoupling the brain from the hands

Anthropic launched Managed Agents, a hosted service that runs long-horizon agents through a small set of durable interfaces. The design virtualizes three components—session (append-only event log), harness (the loop that calls Claude and routes tools), and sandbox (execution environment)—so each can be replaced or fail independently.

Coupling everything in one container created a “pet” that could not be lost and was hard to debug. Decoupling moved the harness out of the container; it now calls the sandbox via execute(name, input) → string. Session state lives outside the harness, enabling wake(sessionId) recovery. Credentials are kept unreachable from the sandbox, closing a prompt-injection vector.

The architecture cut p50 TTFT by roughly 60% and p95 by over 90%. One brain can now address many hands, and hands can be passed between brains. The interfaces deliberately avoid assumptions about future harnesses or model capabilities.

“We virtualized the components of an agent... so the implementation of each to be swapped without disturbing the others.”

Claude Blog: New in Claude Managed Agents: self-hosted sandboxes and MCP tunnels

Claude Managed Agents can now execute tools inside a sandbox the customer controls and reach private MCP servers via tunnels. Self-hosted sandboxes (public beta) run on the customer’s infrastructure or with providers such as Cloudflare, Daytona, Modal, or Vercel; the agent loop stays on Anthropic’s side while tool execution and sensitive data remain inside the enterprise perimeter.

MCP tunnels (research preview) let agents call internal databases, private APIs, and knowledge bases without public endpoints—only an outbound connection is required. Workspace admins manage tunnels from the Claude Console.

Early adopters including Amplitude, Clay, and Rogo are already building design, GTM-engineering, and finance analyst agents on the new primitives.

The Takeaway: Treating the brain as a computer yields medical-scale results far faster than the random walk of drug discovery, and a retinal prosthesis is only the first step toward substrate independence.

Max Hodak, founder and CEO of Science (formerly of Neuralink), positions the company as a medical-device firm whose flagship product Prima is a chip implanted under the retina. Paired with laser-projector glasses it bypasses dead photoreceptors and restores form vision. It received European CE marketing approval in July; first commercial sales are weeks away. Trial patients have completed Sudoku puzzles and read books.

Hodak’s core claim is blunt: “The brain very literally, very clearly, plainly is a computer.” Unlike small-molecule discovery, placing electrodes in motor cortex lets a quadriplegic patient play video games within an hour. Science’s pipeline also includes biohybrid neural interfaces and a perfusion program called Vessel, all aimed at making the brain the upgradable, replaceable center of the human experience and ultimately achieving substrate independence.

“If you can get the visual signal, auditory signal, balance, motor, in and out of the brain, that is an end in itself.”

Replit CEO Amjad Masad called the newly announced partnership with OpenAI “long overdue.” He recalled that before Replit joined YC, Paul Graham asked Sam Altman to recruit the team, and shared Sam’s own retelling of their first meeting. Masad also highlighted that the new Free Mode is fast enough to make coding interactive again and that users can already build a ton with it.

Vercel CEO Guillermo Rauch declared “We’re building AWS for agents.” He also congratulated Jarred Sumner on Bun’s latest release, calling its pursuit of simple, fast and open a perfect match for how Vercel sees the world, and teased that version 0.0.5 ships even smaller (fits on two floppy disks when xz-compressed) together with the team’s most-requested feature.

FPV Ventures partner Nikunj Kothari laid out the new venture math: Anthropic, OpenAI, SpaceX and Cursor have produced trillion- or multi-tens-of-billions outcomes so quickly that LPs are pouring capital into mega-funds. At that scale every investment must be underwritten to a potential trillion-dollar outcome; the cost of omission now dwarfs the cost of commission. Once a deal has heat, entry price barely matters. That is why ambition has become non-negotiable in every pitch.

Claude Code lead Boris Cherny announced that Mythos-class models require additional safety measures; enterprises will own and control their own data and Anthropic retains none. The capability is coming this fall. Thariq added that new Fable safeguards let enterprises keep data and access control on their own infrastructure; the system has been co-developed with roughly 100 companies and will roll out more broadly in the fall.

OpenAI Codex & ChatGPT lead Thibault Sottiaux clarified that usage limits are never changed without transparent community communication. Many reported discrepancies came from users converting subscriptions into API traffic via sub2api for resale or multi-user sharing—behavior flagged by fraud systems. Legitimate use through Sign in with ChatGPT or supported OSS clients is unaffected. He also shipped transparent-image generation in GPT-Image-2 for both ChatGPT and the API.

Meta AI Sr Director Madhu Guru (ex-Google Gemini/Veo) argues the root cause of enterprise AI struggles is the absence of an eval strategy. He recommends a ladder: hill-climb evals that continually push product quality and feature scope; regression evals that catch breakage while climbing; smoke-test evals for non-negotiable safety and basics; and launch evals that approximate real traffic before release.

Practical AI educator Peter Yang shared a gut instinct: much AI output can be improved by a simple loop in which a manager agent “negs” the worker with prompts such as “Are you sure this is the best you can do?” or “Give me 11/10 output.” He also marked crossing 100 k YouTube subscribers and previewed upcoming interviews with AI-evals experts, Replit’s product-engineering lead, OpenAI’s Finance product lead, and the SpaceX Grok bot team.

Box CEO Aaron Levie highlighted a post-training approach that co-optimizes cost and quality: once a company understands a domain deeply and has volume on similar tasks, purpose-built models become viable. Reward shaping that prefers trajectories reducing inference tokens while holding performance constant lets teams gain accuracy without raising cost—an increasingly attractive path for applied-AI companies with vertical expertise.

Latent Space host Swyx celebrated the latest podcast episode co-hosted with Vibhu featuring the founder whose model factory NVIDIA just acquired for $6 B—models that actually beat Thinky on the numbers. He also noted that Matt Pocock’s /grill-me skill has reached the highest echelons (including variations seen by Satya Nadella) and that the new /wayfinder skill plus Ric Mac’s course expand the AI Engineer skills coverage.

Y Combinator President & CEO Garry Tan observed that YC is becoming “the YC for AI Researchers” and, separately, “the YC for consumer hardware.” He also flagged an amusing violation of the “dogfood your own product” rule.

SPC General Partner Aditya Agarwal (ex-Dropbox CTO) stated that the attribute they prize most is Clarity: the ability to be honest, cut through the fog of war, and chart a path through murky water. He held up Sridhar Ramaswamy—who scaled Google Ads from $1 B to $100 B and now leads Snowflake through the AI shift—as the extreme example, and released the latest Minus One episode with him.

Builder Zara Zhang recounted a conversation with Claude while feeling unmotivated. Claude’s line—“Motivation follows action more than it precedes it”—completely reframed how she thinks about motivation.