Claude Managed Agents¶

What they are¶

Claude Managed Agents is a product mode on the Anthropic Claude Platform designed for autonomous, long-running agentic work that does not require a human in the loop at each step. Announced in public beta on May 5, 2026 alongside the financial services agent templates launch, Managed Agents provides the infrastructure for "programmatic use" of Claude at the task level: sessions that can run for hours, tools and credentials scoped per-permission, and a full audit log that compliance and engineering teams can inspect (Anthropic, May 5, 2026).

The core distinction is deployment mode. The same agent templates — the same skills, connectors, and subagents — can be deployed in two ways: (1) as a plugin in Claude Cowork or Claude Code, where Claude works alongside the human on the software already on their desktop (the "alongside" mode), and (2) as a Claude Managed Agent on the Claude Platform, where the template runs autonomously for work that spans a full book of deals or a nightly schedule (the "autonomous" mode). Managed Agents is the second deployment target.

Architecture and building blocks¶

Anthropic describes Managed Agents as pre-assembled with the infrastructure "a firm would otherwise engineer themselves." The four key components are:

Long-running sessions. Unlike a typical API conversation turn, a Managed Agent session can span hours — the example given is a multi-hour month-end close process. The session persists tool call state and model context across the duration of the task without requiring the client to manage continuity manually.

Per-tool permissions. Each tool in the agent's toolbox is gated by a granular permission policy. The agent can only call tools it has been explicitly granted access to, scoped to the specific resources that policy covers. This is the Managed Agents equivalent of the allowlist pattern common across all agentic platforms — but managed at the platform level rather than requiring the developer to implement it in their agent loop.

Managed credential vaults. The platform manages credentials (API keys, OAuth tokens, data-provider access tokens) on behalf of the agent. The agent receives scoped, rotated credentials at runtime rather than holding long-lived secrets in its system prompt. This parallels the "scoped token" pattern used by cloud agents at other vendors — GitHub's per-session tokens, OpenAI's sandboxed Codex sessions — applied here to the full range of enterprise tool integrations an operational agent may need.

Claude Console audit log. Every tool call, every decision, and every intermediate output produced by the Managed Agent is recorded in the Claude Console. Compliance teams and engineering teams can inspect the full execution trace after the fact — or in near-real-time for active sessions. This is a hard requirement for enterprise deployment in regulated industries (financial services, insurance, healthcare) and signals that Managed Agents is designed from the start for enterprise audit requirements rather than treating observability as an afterthought.

Agent templates and cookbooks¶

Anthropic's agent template system supports three deployment targets: Claude Cowork plugin, Claude Code plugin, and Claude Managed Agent cookbook. The same template can be packaged as all three, with the deployment target determining the execution environment and human-in-the-loop model.

The word cookbook is significant: it signals that Anthropic packages the infrastructure choices (session persistence, permission scoping, credential management, audit logging) into a reference implementation that teams can deploy and adapt without building the plumbing from scratch.

Each template bundles three things: - Skills — instructions and domain knowledge for the task. - Connectors — governed access to the data sources the task runs on. - Subagents — additional Claude models called by the main agent for specific sub-tasks (e.g. comparables selection, methodology checks).

The ten financial services templates released at launch (pitch builder, meeting preparer, earnings reviewer, model builder, market researcher, valuation reviewer, general ledger reconciler, month-end closer, statement auditor, KYC screener) are the initial catalogue. These are available at the financial services marketplace and can be adapted to an organisation's own modelling conventions, risk policies, and approval flows.

Relationship to the "alongside" mode¶

In Cowork or Claude Code plugin mode, Claude works alongside the user. The agent has access to desktop software — it can draft emails in Outlook, build models in Excel, write code in the IDE — but the user reviews and approves before anything is sent, filed, or committed. This is the "alongside" mode.

In Managed Agent mode, the same template runs autonomously on the Claude Platform. There is no desktop; the agent uses connectors for data access and managed credentials for tool access. Outputs are staged for review rather than applied immediately, but the session can run for hours without human presence. This is the "autonomous" mode for work "that spans a whole book of deals or a nightly schedule."

The two modes are complementary. For a financial analyst, the M365 add-in is used during the day when the analyst is present and reviewing each step; the Managed Agent runs overnight or across a large batch of files while the analyst is away. Both draw on the same underlying templates and domain knowledge.

The brain/hands architecture¶

Industry observers (MSN, May 2026) have characterised Claude Managed Agents as implementing a "brain/hands" decoupling. The Claude Platform serves as the brain — it reasons, plans, selects tools, and manages the session state. The Cowork application (in alongside mode) or the Managed Agent session (in autonomous mode) serve as the hands — executing the actual tool calls: file edits, email drafts, model builds, data queries. In alongside mode, a human approves each consequential action synchronously. In autonomous mode, the audit log provides post-hoc accountability rather than synchronous approval.

This architecture aligns with the broader pattern visible across the agentic AI category in 2026 — Antigravity's Manager Surface watching parallel agents, Codex Cloud running tasks in ephemeral VMs — but Anthropic's implementation emphasises compliance-grade audit trails and managed credential infrastructure as first-class product features rather than customer-managed components.

Availability¶

Claude Managed Agents launched in public beta on May 5, 2026, on the Claude Platform. Available to organisations on paid plans. The "programmatic use" framing implies Team and Enterprise as the effective deployment tier; individual Pro/Max users access the alongside mode through Cowork plugins rather than Managed Agents directly. Cloud provider routing (Microsoft Azure AI Foundry, Amazon Bedrock, Google Cloud Vertex AI) is supported.

Current limitations¶

Public beta. Platform APIs and cookbook schema may change before GA.
Human review still required for final outputs. Even in autonomous mode, Anthropic's documentation states that users "stay firmly in the loop — reviewing, iterating on, and approving Claude's work before it goes to a client, gets filed, or is acted on." The agent is autonomous in intermediate steps; its final outputs are staged for sign-off.
No published SLA or uptime commitments for the beta platform.
Cookbook catalogue is finance-focused at launch. Broader vertical coverage has not been publicly roadmapped.
No published third-party benchmark validation. The Vals AI Finance Agent benchmark figure (Opus 4.7 at 64.37%) is the primary public benchmark; independent production-scale validation is not yet available.

Sources¶

Changelog¶

2026-05-16 — Page created from Anthropic finance-agents announcement + MSN analysis (Type B/C, confidence 75 moderate). Resolves ALERTS.md claude-agents-upload-blocked item for Managed Agents content. Related pages: claude-agents (parent), claude-agents-m365 (parallel deployment surface).
2026-05-18 — Backlink: added related_pages a2a-protocol, azure-foundry-hosted-agents (auto-sweep; both new pages reference this page as context for cloud-hosted agent infrastructure).