Managed Agents moves orchestration into the infrastructure layer

Anthropic introduced the architecture behind Claude Managed Agents on April 8, 2026, describing a hosted service for long-horizon work built around three separable components: a session log, the agent harness, and an execution sandbox. The company’s argument was that model improvements can make yesterday’s orchestration tricks obsolete. Stable interfaces would let the implementation change without forcing every application to rebuild its surrounding workflow. Anthropic Managed Agents architecture.

Delegation creates a new kind of responsibility: deciding when the delegated work should stop. An autonomous manager can otherwise keep requesting revisions, duplicate tasks, or wait for an output that cannot arrive. A complete multi-agent design therefore needs budgets, cancellation, and a recovery plan as well as an attractive division of labor.

MetaGPT's August 1, 2023, paper used role-based procedures to organize collaboration and address cascading inconsistencies. It provides a research example of making workflow structure explicit. That structure should be assessed alongside the models, because it influences both the errors produced and the opportunities to catch them. MetaGPT paper.

Bound the conversation

In our analysis, each subtask should have a completion condition, a resource budget, and an escalation condition. If two agents repeatedly disagree about missing information, the manager should obtain that information or report the gap. Another round of discussion should need a reason. Stopping is part of competent autonomy, not an admission that coordination has failed.

CAMEL, introduced March 31, 2023, examined autonomous cooperation through role-playing agents. Its research helps frame the question of how communication shapes task completion. It does not remove the application's need to define how conversations terminate or what happens to partially finished work. CAMEL paper.

Recover from the middle

We would test a coordinator by interrupting one worker, invalidating an assumption, and making a dependency unavailable. The system should preserve completed artifacts, cancel obsolete work, and identify what remains. This makes delegation useful under realistic conditions. The manager's most important output may be an accurate account of unfinished work rather than a polished declaration that every agent has completed its role.

The model change that made a workaround stale

Anthropic gave a concrete example: a harness workaround for Sonnet 4.5’s tendency to finish prematurely near its context limit was unnecessary with Opus 4.5. Keeping the workaround would preserve complexity after the original problem had changed. The company therefore framed sessions as append-only histories and sandboxes as execution environments separate from the loop that calls the model. This is an infrastructure design, not evidence that long-running agents no longer make mistakes. Anthropic Managed Agents architecture.

A manager has to manage dependencies

The distinction matters because an orchestration agent is often presented as a model that simply tells other models what to do. Real coordination includes durable state, retries, resource allocation, and deciding when one result is ready for another worker. Some of that work benefits from language-model judgment. Some is better represented by explicit software mechanisms that behave the same way each time.

For example, deciding how to investigate an unfamiliar question can be open-ended. Recording whether a file was successfully saved should not be. A system that asks the model to infer both from conversational history can blur plans with completed actions. Separating the session record from the model’s current context gives the surrounding application a clearer account of what actually happened.

Research agents show the cost of dependency chains

Anthropic’s June 2025 Research account said its lead agent waited for a batch of subagents to finish before proceeding. The arrangement simplified coordination but could leave the whole system waiting for one slow search. The company discussed asynchronous execution as a possible improvement while acknowledging the added difficulty of state consistency and result coordination. That trade-off remains relevant when a managed service hides the implementation behind a simpler interface. Anthropic multi-agent research system.

The business changes when orchestration is hosted

A hosted agent service can remove work from application developers, but it also changes what they must understand. They may no longer operate every worker directly, yet they still need to know which state survives an interruption, which actions can repeat, and how a result is associated with the correct task. The service boundary does not erase these questions; it turns them into part of the product contract.

This creates a different kind of competition from model benchmarks. Two services can call equally capable models while differing in recovery, visibility, and integration. A task that completes in a demonstration may fail in production because an external service is unavailable or a process restarts. The quality of the orchestration layer determines whether the system resumes useful work, duplicates an action, or loses the thread.

The Managed Agents announcement is therefore significant for what it chooses to stabilize. It treats the session and execution environment as durable concepts while allowing the model-facing loop to evolve. That is a plausible response to a fast-changing model frontier. It also suggests that the most useful “manager” in a multi-agent system may be a combination of model judgment and conventional infrastructure. The model allocates attention; the infrastructure preserves the consequences. Judging either in isolation misses how long-running autonomy actually works.

The research behind the release

MetaGPT’s August 2023 paper incorporated software-development procedures and role-based outputs into a multi-agent framework. Intermediate artifacts were intended to organize work and reduce the problems of free-form conversation. The later move toward hosted agent infrastructure addresses a neighboring layer: keeping the process and its state alive while the underlying implementation changes. A development procedure can describe what should happen next, but it cannot by itself guarantee that a worker’s result was stored or that a retry will not duplicate an action. Procedural organization and durable execution solve different parts of the same coordination problem. Original research paper.