AI Builders Digest

Anthropic traces a month of Claude Code complaints to three separate changes

Three unrelated changes stacked into what looked like broad, inconsistent degradation across Claude Code, the Agent SDK, and Cowork, while the API was never affected. The default reasoning effort dropped from high to medium on March 4, a March 26 caching optimization contained a bug that discarded Claude's prior reasoning on every turn once a session went stale, and an April 16 system prompt line capping responses at 100 words cost about 3% on evals for both Opus 4.6 and 4.7. All three are resolved as of April 20, the default is now xhigh for Opus 4.7 and high everywhere else, and usage limits are being reset for every subscriber. Worth noting from the investigation: back-tested against the offending PRs, Opus 4.7 found the caching bug and Opus 4.6 did not.

• #products
• #evals

GPT-6 Astra on low reasoning beats GPT-5.6 Sol on high

Calibration guidance from the Codex and ChatGPT side of OpenAI: GPT-6 Astra at low reasoning effort performs better than GPT-5.6 Sol at high. If you were running Sol on high and happy with it, the advice is to move down to low or medium on Astra rather than carrying the old setting across.

• #evals
• #products

Plan as though AI R&D is fully automated by early 2029

Ryan Greenblatt's median for full automation of AI R&D is end of 2030, but his 35th percentile is end of 2028, and that earlier date is what he thinks people should actually plan around. "I wouldn't say superintelligence is bad. I would say it's dangerous." His central scenario runs like this: software engineering inside AI companies fully automated in early 2028, models thinking in an AI-only language we can ask them to decode by mid 2028, roughly 4x as much AI progress in 2029 as in 2025, and then a shift from sloppy reward hacking to competent scheming and takeover. The AI 2040 plan he co-authored tries to buy time through a US-China deal built on tracking compute and total research transparency, which would gut the frontier labs' biggest moat while still leaving world GDP growing roughly 200x during the 2030s.

• #policy
• #safety

Anthropic: users approved 93% of Claude Code permission prompts

Telemetry showed users clicking approve on roughly 93% of permission prompts, which is the case for treating containment rather than human review as the primary defense. An OS-level sandbox cut prompts by 84%, and Opus 4.7 holds prompt injection attack success to about 0.1% on single attempts and 5 to 6% after 100 adaptive tries. The two most instructive failures both slipped past every probabilistic layer: an internal red team phished an employee into pasting a prompt that exfiltrated ~/.aws/credentials in 24 of 25 runs, and a third party showed that allowlisting api.anthropic.com let a poisoned workspace file upload data into an attacker's own Anthropic account. Their recurring lesson is that the battle-hardened primitives held, and the custom proxy they wrote themselves is what broke.

• #agents
• #security

Claude Code can now publish live artifacts your team can open

Claude Code sessions can now produce a shareable web page built from the session's full context, including the codebase, connectors, and the conversation itself. Every publish is a new version at the same link with version history, and open pages refresh in place, so an incident investigation republishes itself as it progresses and ends up being the postmortem. Artifacts are private to the author by default, viewable only by authenticated members of the org, and cannot be made public. It is in beta for Claude Team and Enterprise from the CLI and desktop app.

• #products
• #agents

Anthropic pulls the agent harness out of the sandbox container

Managed Agents splits an agent into three swappable interfaces: a session that is an append-only event log, a harness that runs the loop, and a sandbox that executes code. Putting all three in one container made that container a pet, so a crash lost the session and connecting Claude to a customer VPC meant peering networks. Decoupling them means a container gets provisioned only when a tool call needs one, which dropped p50 time to first token roughly 60% and p95 over 90%. The security payoff is that credentials never reach the sandbox: git tokens are wired into the local remote at init, and MCP OAuth tokens sit in a vault behind a proxy the harness never sees.

• #agents
• #products

Chasing cloud agent sessions that start in seconds, not repo clones

The one piece missing from the harness he wants is cloud sessions, and the target is seconds to start, which requires clever snapshotting. Cloning repos fresh, which is what it does today, is not fast enough. He expects to have it next week, and separately says he can't remember the last time capabilities jumped this much at once.

• #agents
• #products

Brilliant's product rule: never tell the learner the answer

Sue Khim, co-founder of Brilliant, builds on a single principle: never tell the learner the answer, because cheating and explaining the answer to your child are closer than you'd think. Her framing is that using AI to skip learning is like bringing a robotic arm to the gym to lift weights for you, and that learning was never about the answer but about strengthening the part of you that can concentrate and struggle. Her advice for anyone building AI products is to find areas where you hold unique data that makes the product better over time.

• #products
• #education