Claude Managed Agents Just Killed Openclaw

The discussion revolves around "Managed Agents" introduced by Claude, presented as a robust and efficient way to develop and deploy AI agents. The speaker expresses enthusiasm for the concept, particularly the idea of agents researching acquisition targets and building investment theses, noting its practical application. Managed Agents aim to streamline the process of shipping production-grade agents by handling infrastructure complexities such as secure sandboxing, checkpointing, credential management, and tracing. This allows developers to focus on user experience rather than operational overhead, shortening the development cycle from months to days. Early customer examples highlight the utility of Managed Agents. Notion, for instance, uses them to create an agent orchestration platform, facilitating human and agent collaboration. A product manager at Notion described how Managed Agents enabled rapid prototyping, with Claude running autonomously to generate prototypes, significantly accelerating their workflow. Notion leverages Claude's ability to handle complex, long-running tasks, integrating it via an API to manage client onboarding workflows and provide full context from client databases and task boards. The speaker then delves into the technical details and implications of Managed Agents. They are described as a suite of composable APIs for building and deploying cloud-hosted agents at scale. The system includes a performance-tuned agent harness, production infrastructure, and built-in orchestration that manages tool calls, context, and error recovery. Key features include production-grade security, long-running autonomous sessions, multi-agent coordination for parallelizing complex work, and trusted governance with scoped permissions and identity management. Claude models are specifically designed for agentic work, and Managed Agents are purpose-built to maximize their effectiveness. Users define outcomes and success criteria, and Claude self-evaluates and iterates to achieve them. This functionality is currently available in research preview. Internal testing has shown that Managed Agents improve outcome task success by up to 10 points over standard prompting loops for structured file generation. Managed Agents are already being used across various production use cases, including coding agents that read codebases and plan fixes, productivity agents that join projects and deliver work, and finance/legal agents that process documents. These applications demonstrate faster time-to-value for users. Notion, for example, allows teams to delegate work to Claude directly within their workspace, enabling parallel execution of numerous tasks for engineers and knowledge workers alike. Rakuten and Asana are also mentioned as users building AI teammates with collaborative AI agents. Regarding pricing, Managed Agents are consumption-based, with standard Claude platform rates applying, plus an additional $0.08 per session hour for active runtime. The speaker clarifies that this $0.08 is for the "rent" or runtime of the agent, with the primary cost still being the token usage. The speaker explores how to get started with Managed Agents, noting their availability on the Claude Platform, accessible via documentation, the Claude Console, or the CLI. Developers can use the latest version of Claude Code and its built-in Claude API skill. The speaker initially struggles to find clear instructions but eventually learns that within Claude Code, typing a specific command will initiate the onboarding process for Managed Agents via the Claude Agent SDK. A key distinction is made regarding billing: Managed Agents built via the Agent SDK are billed through the user's Claude Code subscription plan, not their Anthropic API key. This means they utilize the Claude Code usage quota and are separate from direct Anthropic API usage. The Claude API skill, when loaded, serves as documentation and does not enable billing or API connections. The speaker engages in an interactive session with an AI, asking it to create useful agents based on their existing data and bots. The AI proposes several agent ideas: a "Strategy Analyzer" to read trade CSVs and calculate win rates, a "Market Scanner" for new opportunities on Polymarket, a "Code Reviewer" for bot bugs, and a "Research Agent" for news relevant to positions. The AI suggests building the "Strategy Analyzer" first due to its immediate utility. The AI further explains that Managed Agents differ from regular API calls by looping autonomously, using tools, evaluating results, and iterating, much like a human solving a problem. They can also spawn sub-agents for specialized tasks, akin to a project manager delegating to analysts. The AI then creates three agents: a Cloud Code Reviewer, a Strategy Analyzer, and a Market Scanner. The speaker attempts to run these agents but initially sees no output, questioning if they are working. The AI clarifies that the agents are running, but the interim work is invisible, with only the final results being displayed. The speaker then observes the agents actively running. The conversation shifts to comparing Managed Agents with OpenClaw, an open-source AI agent. The AI clarifies that Anthropic runs Managed Agents in the cloud, while OpenClaw is run locally by the user. Managed Agents offer fully managed outcomes, while OpenClaw is self-managed. The speaker also notes that Managed Agents were released today, hence the initial excitement. The speaker encounters some confusion regarding subscription models and potential scams, clarifying their own product offerings (lifetime access and $5 zooms, not subscriptions or exchange services). They also address an old product page that might have caused confusion about pricing. Further attempts to use Managed Agents lead to more questions about their execution, particularly whether they run in separate terminals or if the built files are indeed the new Managed Agents. The AI confirms that the agents built through the SDK run on the Claude Code plan, not the API key, and offers to build a Managed Agents version if the user provides their API key. The speaker decides to monitor what others develop, acknowledging the complexity and potential need for expert guidance.