Claude Managed Agents
Claude Managed Agents is a managed agent infrastructure service provided by Anthropic for deploying and running autonomous agent tasks based on Claude. The service supports self-hosted sandbox environments for secure, isolated code execution and uses MCP (Model Context Protocol) tunneling to enable standardized connections between agents and external tools and services, targeting developers and enterprises that need to run complex multi-step automation tasks in controlled environments.
Timeline (last 90 days)
For clients integrating a managed agent, the pragmatic path is for the client to first perform data minimization and anonymization, uploading only aggregated features rather than raw URL lists, and treating Managed Agents as a generation layer rather than a data layer.
ai-newtab is an open-source project that, through a Chrome extension, sends the user's browsing history to a Claude Managed Agents session to generate a daily-updated personalized new tab page.
Compared with the traditional bare API-calling model, the managed agent model abstracts the agent lifecycle into a black box on the platform side, with the client only needing to submit input and receive output, but at the cost of partially uncontrollable runtime behavior and the requirement that data be processed through the platform.
Under the mechanism in which large models bill by input and output tokens, resending the full browsing history every day causes the input token cost to grow linearly with the number of days of use; having the client perform deduplication, aggregation, or incremental summarization can keep the input volume at a roughly constant level.
The length of state maintained in a session affects cost: a long session carries prior context in each round, while a short session with cold-start input is smaller but may sacrifice continuity, and both of these cost levers are held on the client side.
ai-newtab's teardown points out that when integrating a managed agent into a client, one must face design decisions such as data boundaries, permission models, and output safety, which are not defects of Managed Agents itself but the developer's responsibility.
In Anthropic's tests, a codebase of 116,000 lines of code had 70 bugs planted in it; a single agent found 14, 15, and 27 bugs across three rounds respectively, while a multi-agent workflow consistently found 66 bugs across all three rounds.
Claude Managed Agents dynamic workflows are reportedly capable of orchestrating up to 1,000 agents in a single run.
Anthropic announced a series of new feature updates for Claude Managed Agents at the Code with Claude event.
Outcomes, multi-agent orchestration, and Webhooks have entered public beta.
1 more timeline events
All Facts (15)
Claude Managed Agents的代理循环（Agent Loop）仍运行在Anthropic基础设施上，而工具执行迁移到企业自己配置的环境中
95%Unverified顾问策略（Advisor Strategy）是一种多Agent协作模式，允许主Agent在执行任务时咨询其他专业Agent的意见
90%UnverifiedClaude Managed Agents的代码执行能力允许Agent直接编写并执行代码，完成数据分析、数学计算、文件处理等任务
85%UnverifiedAnthropic承诺所有新功能在原生Claude API上线的同一天也会同步部署到Managed Agents服务中
85%UnverifiedFor clients integrating a managed agent, the pragmatic path is for the client to first perform data minimization and anonymization, uploading only aggregated features rather than raw URL lists, and treating Managed Agents as a generation layer rather than a data layer.
50%Unverifiedai-newtab is an open-source project that, through a Chrome extension, sends the user's browsing history to a Claude Managed Agents session to generate a daily-updated personalized new tab page.
50%UnverifiedCompared with the traditional bare API-calling model, the managed agent model abstracts the agent lifecycle into a black box on the platform side, with the client only needing to submit input and receive output, but at the cost of partially uncontrollable runtime behavior and the requirement that data be processed through the platform.
50%UnverifiedUnder the mechanism in which large models bill by input and output tokens, resending the full browsing history every day causes the input token cost to grow linearly with the number of days of use; having the client perform deduplication, aggregation, or incremental summarization can keep the input volume at a roughly constant level.
50%UnverifiedThe length of state maintained in a session affects cost: a long session carries prior context in each round, while a short session with cold-start input is smaller but may sacrifice continuity, and both of these cost levers are held on the client side.
50%Unverifiedai-newtab's teardown points out that when integrating a managed agent into a client, one must face design decisions such as data boundaries, permission models, and output safety, which are not defects of Managed Agents itself but the developer's responsibility.
50%UnverifiedIn Anthropic's tests, a codebase of 116,000 lines of code had 70 bugs planted in it; a single agent found 14, 15, and 27 bugs across three rounds respectively, while a multi-agent workflow consistently found 66 bugs across all three rounds.
50%UnverifiedClaude Managed Agents dynamic workflows are reportedly capable of orchestrating up to 1,000 agents in a single run.
50%UnverifiedAn Agent's ability to self-verify determines the upper limit of the number of Agents a developer can manage in parallel; if Agents lack self-checking capabilities, human intervention becomes a bottleneck for parallel scaling.
50%UnverifiedAnthropic announced a series of new feature updates for Claude Managed Agents at the Code with Claude event.
50%UnverifiedOutcomes, multi-agent orchestration, and Webhooks have entered public beta.
50%