Lead: Anthropic says recent complaints about Claude Code quality came from three separate product changes—not a model or API degradation—and all have now been fixed as of April 20 (v2.1.116).

Numbers:

• 3 distinct issues affected Claude Code, the Claude Agent SDK, and Claude Cowork
• March 4: default reasoning effort changed from high to medium, then reverted April 7
• March 26: a caching bug was fixed April 10 in v2.1.101
• April 16: a verbosity-reducing prompt change was reverted April 20
• One eval showed a 3% drop for Opus 4.6 and 4.7 after prompt ablations

So What: For builders, the key takeaway is that product-layer changes can look like model regressions even when the API is fine. Anthropic is now resetting usage limits for all subscribers and tightening release controls: broader per-model evals, more ablations, soak periods, gradual rollouts, and better internal dogfooding on the exact public build. The company also plans to improve its code review tooling and add guidance so model-specific changes are gated to the right model. As Anthropic put it, “We never intentionally degrade our models,” and the fixes were driven largely by reproducible user feedback.

Lead: Anthropic introduced Managed Agents, a hosted Claude Platform service for long-horizon agents that separates the “brain” (Claude and its harness) from the “hands” (sandboxes/tools) and the durable session log.

Numbers:

• p50 time-to-first-token dropped roughly 60%.
• p95 time-to-first-token dropped over 90%.
• The architecture supports many stateless harnesses and many independent tool environments.

So What: The big shift is architectural: sessions live outside the harness, tools are invoked through a simple `execute(name, input) → string` interface, and credentials stay out of the sandbox. That makes agents easier to recover, safer against prompt injection, and more flexible across VPCs, Git, MCP, and custom tools. Anthropic’s message is that harness assumptions can go stale as models improve, so Managed Agents is designed as a “meta-harness” that can survive future changes. As the post puts it, “we aimed to design a system for ‘programs as yet unthought of.’” For builders, the practical takeaway is to move long-running state into durable sessions, treat compute and tools as swappable interfaces, and only provision sandboxes when needed.

Lead: Claude Managed Agents can now run tool execution inside self-hosted sandboxes and connect to private MCP servers, keeping agent work inside an enterprise-controlled perimeter.

Numbers:

• Self-hosted sandboxes are in public beta on the Claude Platform.
• MCP tunnels are in research preview and require access.
• Supported sandbox providers include Cloudflare, Daytona, Modal, and Vercel.

So What: This is a meaningful step for teams that want Claude agents to work on sensitive code, files, and internal systems without sending data outside their environment. Anthropic says the “agent loop” stays on its infrastructure, while execution moves to your chosen sandbox, where you control compute, runtime image, logging, and network policy. For private tools, MCP tunnels let agents reach internal databases, APIs, knowledge bases, and ticketing systems through a single outbound connection—“no inbound firewall rules, no public endpoints.” Builders can use this to ship more capable enterprise agents with tighter security, better observability, and infrastructure that matches workload needs, including long-running or GPU-heavy jobs.