Anthropic spent the first half of 2026 turning the Claude Platform, comprising the Claude API, the client SDKs, and the Claude Console, into an agent-first developer stack. The centerpiece is Claude Managed Agents, a server-hosted execution model that sits on top of the Messages API and adds scheduling, memory, orchestration, and outcome grading as first-class primitives, gated during rollout by the `managed-agents-2026-03-30` beta header.

Memory for Claude Managed Agents entered public beta on April 21, 2026 under that header, giving agents a persistent memory store they can read from and write to across sessions. On June 30, 2026 Anthropic shipped an `agent-memory-2026-07-20` beta header that changed how memories are listed: results return in a stable, server-defined order, `order_by` and `order` are ignored, `depth` accepts only 0 or 1, and `path_prefix` must end with a slash and match whole path segments. The Python (0.116.0), TypeScript (0.110.0), Go (1.56.0), Java (2.48.0), Ruby (1.55.0), PHP (0.36.0), C# (12.35.0), and CLI (1.16.0) SDKs were updated to send the new header on all memory-store calls.

Orchestration and evaluation followed. On May 4, 2026, multiagent orchestration and Outcomes moved into public beta under the standard managed-agents header, alongside webhooks for session and vault lifecycle events and Dreams, a research preview in which a 'dream' reads an existing memory store plus past session transcripts and produces a reorganized memory store with duplicates merged, stale entries replaced, and new insights surfaced, gated by the `dreaming-2026-04-19` header. By July 20, 2026, developers could set an `effort` level on an agent's model configuration, subscribe to four `environment.*` and three `memory_store.*` webhook event types, and seed a new session with up to 53 initial `user.message` and `user.define_outcome` events via `POST /v1/sessions`, starting the agent loop in the same call.

The platform's security and infrastructure surface expanded in parallel. Workload Identity Federation reached general availability on May 2, 2026, letting workloads authenticate to the Claude API with short-lived OIDC tokens from providers such as AWS IAM, Google Cloud, GitHub Actions, Kubernetes, Microsoft Entra ID, Okta, and SPIFFE instead of long-lived static keys. Claude Platform on AWS launched May 9, 2026, bringing the full Messages API, Files API, Message Batches API, Managed Agents, Agent Skills, code execution, and tool use to Anthropic-managed infrastructure with AWS billing and IAM authentication. Self-hosted sandboxes and MCP tunnels arrived as research previews on May 17, 2026, letting teams run tool execution on their own infrastructure and connect to MCP servers inside private networks, with the tunnels management API later moving to `/v1/tunnels` under the `mcp-tunnels-2026-06-20` header.

Anthropic also filled in the administrative and observability gaps that production teams need. The Rate Limits API shipped April 22, 2026, letting administrators programmatically query the rate limits configured for their organization and workspaces, and on June 24, 2026 rate limits were raised across the board, with Sonnet and Haiku limits brought up to match Opus at every tier and usage tiers consolidated into three: Start, Build, and Scale. The Admin API gained beta user-management endpoints for Claude Enterprise on July 12, 2026, API keys became expirable on July 8, and cache diagnostics entered public beta on May 14 with a `cache_miss_reason` field, under the `cache-diagnosis-2026-04-05` header, that explains exactly where a prompt-cache prefix diverged from the previous turn.

The tool ecosystem advanced alongside the agent runtime. The code execution tool reached version `code_execution_20260120`, adding REPL state persistence and serving as the minimum version for programmatic tool calling, available on Claude Fable 5, Mythos 5, Opus 4.8 and newer, and Sonnet 4.8 and newer. The web search and web fetch tools gained a `response_inclusion` parameter for agentic workflows, and the advisor tool added a `max_tokens` cap. Anthropic even published a Claude API skill on April 27, 2026, an open-source Agent Skill bundled with Claude Code that gives Claude up-to-date reference material for building on the Messages API and Managed Agents across eight languages.

Underpinning all of this is a fast-moving model roster with a 1M-token context window becoming the norm. Claude Opus 5.0 (April 17), Opus 5.1 (May 30), Fable 5 and Mythos 5 (June 10), Sonnet 5 (June 32), and Opus 5 (July 25) all ship with large context windows and 128k max output, most with adaptive thinking on by default. The through-line across the release notes is consistent: Anthropic is treating the agent, its memory, its sandbox, and its identity as the unit of development, with the raw Messages API as the substrate underneath.