Anthropic on May 6 opened its annual Code w/ Claude developer conference in San Francisco with a coordinated set of releases aimed at the production-agent layer that sits on top of the Claude API. The conference week is the most consequential single drop for the coding-models space in the past twelve months.

Headline announcements:

### Claude Managed Agents

A hosted offering on top of the Claude API that bundles four capabilities: multiagent orchestration, scheduled "dreaming" runs that let agents iterate on tasks during idle time, Outcomes telemetry for tracking agent execution, and webhook integration. For coding-models operators, the practical change is that the multiagent scaffolding for things like "code-writing agent + code-review agent + test-running agent" no longer has to live in self-managed infrastructure. The orchestration moves into Anthropic's plane.

### Higher API rate limits

Claude Code and the Opus tier both received rate-limit increases on May 6. The lab framed this as a recognition that the bottleneck for production agentic coding is no longer model availability but the scaffolding around it.

### Self-hosted sandboxes (public beta)

For coding workloads that touch sensitive code, packages, or internal tooling, the new self-hosted sandbox option keeps execution inside the operator's own infrastructure while letting the Claude orchestration plane manage the agent. The public beta tag is meaningful — the option is operable today but the contract isn't final-form yet.

### MCP tunnels (research preview)

A way to expose tools running on an MCP server inside a private network to a Claude client outside it, without opening inbound ports. For coding deployments that need to access internal tooling — package managers, build systems, internal CLIs — without permissive networking, this is the new operator-friendly path.

### What we don't have from Code w/ Claude

Public benchmark scores for Claude Code on SWE-bench Verified, OSWorld-Agent, or TAU-bench at the new rate limits and orchestration layer.
A side-by-side comparison of Claude Managed Agents versus self-built Claude orchestration on a representative coding workload, with cost and latency.

### The competitive frame

Code w/ Claude lands the same week that Microsoft unveiled Agent 365 (the IT-administration plane for autonomous agents) and Copilot Cowork's iOS/Android expansion. Microsoft and Anthropic are now pushing converging positions: the agent scaffolding is the product, the model is the substrate.

For coding-model operators currently running Claude Code in production, the practical question for the next month is whether to migrate orchestration scaffolding onto Claude Managed Agents or to maintain self-managed control. The right answer depends on workload sensitivity, but the cost of the migration just dropped meaningfully.

Claude Code goes agentic at Code w/ Claude: Managed Agents, higher rate limits, and self-hosted sandboxes

Sources