Technology RadarTechnology Radar

Claude Managed Agents

agentmulti-agentworkflow
Assess

Claude Managed Agents is Anthropic's hosted orchestration layer for production AI agents — launched April 8, 2026 in public beta. It provides sandboxed code execution, checkpointing, credential management, scoped permissions, and end-to-end tracing as managed infrastructure, so teams can go from prototype to production without building those primitives themselves.

This is the direct Anthropic answer to the known limitations of Claude Code Agent Teams — particularly the lack of session resumption, limited observability, and no persistence across disconnections.

What It Is

Claude Managed Agents exposes a suite of composable APIs hosted on the Claude Platform. The built-in orchestration harness handles when to call tools, how to manage context, and how to recover from errors. Agents run in sandboxed environments with persistent memory, allowing them to retain context across interactions and complete multi-step tasks without any user-built orchestration plumbing.

Key capabilities at launch:

Capability What It Provides
Sandboxed execution Isolated runtime per agent; code executes in a contained environment
Checkpointing Agent state persists across sessions; long-running tasks survive disconnections
Credential management Secure, scoped credential injection — no manual secret handling
Scoped permissions Fine-grained tool and resource access controls per agent
End-to-end tracing Full observability into agent decisions, tool calls, and outputs
Long-running sessions Autonomous operation for hours; progress persists even if client disconnects

How It Compares to Claude Code Agent Teams

Claude Code Agent Teams (experimental, v2.1.32+) is the local, developer-facing multi-agent feature. Managed Agents is the cloud-native, API-first production equivalent. They target different scenarios:

Dimension Claude Code Agent Teams Claude Managed Agents
Deployment Local terminal / desktop Cloud-hosted API
Session resumption Not supported Checkpointed and resumable
Observability Hook events only End-to-end tracing built in
Persistence Context window only Persistent memory across sessions
Long-running tasks Bounded by terminal session Hours of autonomous operation
Sandboxing Application-level permissions Managed sandboxed environment
Multi-agent coordination Task list + mailbox (file-based) Public beta (May 2026)
Cost API tokens only API tokens + $0.08/session-hour
Target user Developer, in-project interactive work Team, production integration, API consumers

Memory (Public Beta, April 23, 2026)

Anthropic added persistent cross-session memory to Managed Agents in a public beta two weeks after launch. Memory stores agent knowledge as files on a managed filesystem — accessible via API or the Claude Console. Key properties:

  • Filesystem-backed: memories are files you can export, inspect, and edit directly
  • Scoped permissions: per-agent memory isolation with organization-level governance
  • Audit trail: all memory reads and writes are logged, with rollback and redaction via API
  • Opt-in: agents only access memory when explicitly configured to do so

Early adopters: Netflix, Rakuten, Wisedocs, Ando — Wisedocs reports a 97% reduction in first-pass errors and 30% speed increase in document verification workflows (per Anthropic). Results are from named but unvalidated case studies; independent benchmarks are not yet available.

Dreaming (Research Preview, May 2026)

Announced at Code with Claude 2026 (May 6, 2026), Dreaming is a scheduled background process that reviews an agent's past sessions and memory stores, extracts recurring patterns, and writes learnings as plain-text notes and structured "playbooks" for future sessions. Dreaming does not modify model weights — all improvements live in the memory layer, making them inspectable, reversible, and exportable via the same Memory API. Access requires a separate request (research preview only).

The design analogy: when a person completes a complex workflow, they often distil what they learned into an SOP or runbook. Dreaming automates that capture — after each session, the agent does the same review without any manual human effort. Future sessions reference the playbooks as lightweight context, improving without retraining.

Outcomes (Public Beta, May 2026)

Also announced at Code with Claude 2026, Outcomes adds a grade-and-revise loop to any session. You write a rubric — a markdown document with explicit, checkable criteria — and attach it via user.define_outcome. The harness provisions a separate grader agent (a fresh context window with only the rubric and the artifact) that evaluates the output. When criteria aren't met, the grader pinpoints exactly what needs to change and the writer takes another pass, up to a configurable iteration cap (default 3, max 20).

client.beta.sessions.events.send(
    session.id,
    events=[{
        "type": "user.define_outcome",
        "description": TASK,
        "rubric": {"type": "text", "content": RUBRIC},
        "max_iterations": 5,
    }],
)

Key design principles for effective rubrics: use checkable criteria (exact figures, not vague goals), require evidence (URLs that must fetch, verbatim quotes), and anticipate shortcuts (distinguish an 8-K press release from a 10-K filing, for example — the grader enforces distinctions the task alone wouldn't catch). On Anthropic's internal benchmarks, Outcomes lifted task success by up to 10 percentage points on the hardest problems, with no model change.

Grader result states: satisfied (loop terminates), needs_revision (feedback returned to writer), max_iterations_reached, failed.

Outcomes is a strong fit for attention-to-detail tasks, exhaustive checklists, and citation verification; less useful for purely subjective quality without clear criteria.

Known Limitations (Public Beta)

  • Multi-agent coordination at public beta, not GA — moved from research preview to public beta at Code with Claude 2026 (May 2026); the feature that most directly parallels Agent Teams is now accessible without a separate access request, but production SLAs and GA stability guarantees have not yet been published
  • Dreaming still in research preview — requires a separate access request; cannot be relied on for production workflows yet; no GA timeline announced
  • Priced per session-hour — $0.08/active session-hour on top of standard token rates; cost can accumulate for long-running workflows
  • No self-hosted option — unlike LangGraph or Google ADK, this is cloud-only; regulated industries with data-residency requirements may not qualify

Why Assess (Not Trial)

Managed Agents has shipped at a fast pace: public beta (April 8), Memory (April 23), and Dreaming + Outcomes + multi-agent orchestration public beta (May 6) all within six weeks. Multi-agent coordination moved from research preview to public beta — the main "assess" blocker is now advancing toward GA. But production SLAs for multi-agent orchestration have not yet been published, Dreaming is still research preview, and the pricing model for long-running agents still needs real-world validation beyond vendor-provided case studies (the three early adopter results — Netflix, Rakuten, Wisedocs — come from Anthropic's own announcements). Revisit for Trial once multi-agent orchestration reaches GA with published SLAs and independent cost-of-ownership data is available.

Relationship to Other Orchestrators

  • Claude Code — Agent Teams is the local multi-agent feature; Managed Agents is the cloud complement. They are not mutually exclusive — Agent Teams is for interactive development, Managed Agents for production deployment.
  • LangGraph — Trial. Build-your-own with the highest control and checkpointing maturity; model-agnostic. Managed Agents is a buy option if you want to skip building the infrastructure.
  • Microsoft Agent Framework — Now at GA (April 3, 2026). Also supports checkpointing and pause/resume, model-agnostic, .NET + Python. Better fit for Azure-committed enterprises.
  • Google ADK — Event-driven, model-agnostic orchestration; strong A2A and MCP integration; deploy to Cloud Run or Vertex AI. See the ADK architecture teardown. Managed Agents is more opinionated but less infrastructure work.

Sources