Anthropic Claude Agents: Every Feature Shipped in 2026

Claude Managed Agents shipped on April 8, 2026, and Anthropic has added something almost every month since. Here is the dated record: what landed, when, and the numbers they published with it.

📜 Table of Contents

• The Release Timeline
• What Managed Agents Actually Is
• Six Things That Shipped, and What They Replace
• The Numbers Anthropic Published
• Managed Agents vs. the Agent SDK
• What to Watch
• FAQ

Anthropic shipped Claude Managed Agents on April 8, 2026, and has added something to it almost every month since. This is the dated record of what changed — the capabilities, the numbers Anthropic published with them, and what each one actually replaces in your own stack.

The Release Timeline

What Managed Agents Actually Is

It is three resources, not one product. An agent is a versioned configuration: model, system prompt, tools, MCP servers, skills. An environment is where code runs: the sandbox, its networking rules, its preinstalled packages. A session pairs one agent with one environment and gets its own sandbox instance.

The architectural claim behind all of it is that the harness calling Claude runs separately from the sandbox executing code. Anthropic calls this decoupling the brain from the hands. It has three consequences they measured and published: credentials never sit next to generated code, Claude starts reasoning before a container exists, and every model call and tool result is appended to an event log that survives the process.

Six Things That Shipped, and What They Replace

Memory. Agents learn across sessions from a built-in store. Before this, every session started cold and you re-fed the same context. Memory arrived in beta on 2026-04-23. On 2026-11-02, self-hosted sandboxes gained the same ability: the Python, TypeScript, and Go SDK workers download each attached store into the sandbox at its mount path and sync changes back.

Dreaming. A scheduled process that reads past sessions and memory stores, extracts patterns, and curates what the agent keeps. Without it, memory fills with noise; with it, recurring mistakes get written down as something the agent can act on.

Multiagent orchestration. Three modes, and the documentation is specific about when each applies. Subagents are for work the agent delegates itself and follows up on — delegation is one level deep, capped at 25 child threads per session, idle ones included. Dynamic workflows are for work with many pieces: the agent writes a program that runs many agents in phases and combines results, and the server runs it in the background. Advisor is for one agent doing the work with a stronger model consulted at planning or review, billed at the advisor's rates.

Scheduled deployments. Run sessions on a cron schedule without running your own scheduler. This is the difference between an agent you trigger and an agent that shows up.

Self-hosted sandboxes. The hands can live in your VPC. Code, filesystem, and network egress never leave your environment. Paired with MCP tunnels, which control how Anthropic reaches MCP servers inside your private network, you get to decide exactly what crosses the boundary.

Vaults. Credentials for MCPs, CLIs, and repos live in a separate store, envelope-encrypted, fetched and decrypted by a proxy only on demand. As of 2026-10-09 this includes environment-variable credentials, which is what most CLI tooling expects.

The Numbers Anthropic Published

Time to first token dropped roughly 60% at p50 and over 90% at p95. The cause is specific: without the decoupled architecture, a container spins up for every session, including sessions where the agent only thinks and never runs a tool. With it, Claude begins reasoning while the environment spins up in parallel, and think-only sessions skip the container entirely.

Up to 10 points better task success on structured file generation, measured internally against a standard prompting loop, with the largest gains on the hardest problems.

25 child threads per session, delegation one level deep. That is a real ceiling if you were planning to fan out broadly.

Managed Agents vs. the Agent SDK

They are not competitors, and Anthropic's own documentation is clear about the split. The Agent SDK runs the Claude Code binary inside a process you operate: you get its built-in tools, permissions, sessions, and hooks, and you handle hosting. Managed Agents puts the harness on Anthropic's side and gives you sessions, memory, sandboxes, vaults, and scheduling through the API. If you need the full loop under your control, take the SDK. If the loop is commodity and the infrastructure is your problem, take Managed Agents.

What to Watch

Multiagent coordination and outcome-based self-evaluation are still behind a research preview request form, not generally available. Source-of-truth for anything in this article is the Anthropic release notes, which are dated and public.

FAQ

When did Claude Managed Agents launch?
April 8, 2026, in beta, under the managed-agents-2026-04-01 beta header. The same release introduced the ant CLI.

Is it generally available?
The core is in public beta on the Claude Platform. Multiagent coordination and agent self-evaluation against outcomes are in research preview and require requesting access.

Can agents run on my own infrastructure?
Yes. Self-hosted sandboxes keep tool execution inside your VPC. MCP tunnels connect Anthropic to MCP servers on your private network. Scheduled deployments run sessions on a cron schedule.

Do agents remember between sessions?
Yes, through memory stores, in beta since April 23, 2026, plus a scheduled "dreaming" process that curates what gets kept.

How much faster is it than a standard agent loop?
Anthropic reports about 60% lower time to first token at p50 and over 90% at p95, and up to 10 points better success on structured file generation in internal testing.