Claude agents: billable tools and a wobble

Anthropic’s new Managed Agents can use built-in tools and customer-run server-side tools, and some of those tool calls can add separate usage charges. (platform.claude.com) Managed Agents is Anthropic’s hosted runtime for long-running Claude tasks: developers define the model, prompt, tools and environment, then start sessions through a beta API header dated `managed-agents-2026-04-01`. The quickstart says the default toolset includes bash, file operations and web search. (platform.claude.com) Anthropic’s engineering post describes the product as a split between the “brain,” the “hands” and a durable session log, with Anthropic handling tool execution, context management and recovery when parts fail. That means companies can ship an agent without building their own loop, sandbox and state layer first. (anthropic.com) The pricing wrinkle is that Managed Agents is not billed on one meter. Anthropic’s public materials say developers pay model usage and session runtime, and Anthropic’s documentation also notes that server-side tool usage such as web search may be metered separately on a per-call basis. (platform.claude.com) That cost structure lands at the same moment Anthropic is pushing agents as production infrastructure rather than a demo feature. The company’s launch materials say early users include Notion, Rakuten and Asana, and the pitch is that Anthropic handles the plumbing while customers watch token, runtime and tool bills. (anthropic.com) Then, on April 13, Claude users reported intermittent HTTP 500 errors across claude.ai, the application programming interface and Claude Code. Third-party outage trackers and news reports collected complaints even while some official status views still showed broad service availability. (downdetector.com) (cybersecuritynews.com) Anthropic’s own status page later logged an April 13 incident as “Claude.ai down,” saying elevated login errors affected Claude.ai and Claude Code from 15:31 to 16:19 Coordinated Universal Time. The same page now shows Claude API in “Degraded Performance” on April 14 because of a separate usage and analytics admin endpoint incident. (status.claude.com) That mismatch is familiar in cloud software: a narrow failure can hit logins, one region or one product surface before a status page flips from green to red. For developers evaluating Managed Agents this week, the new question is not only what the agent can do, but which tool calls are billable and which outages show up where. (status.claude.com)