A CTO told me his team spent four months on an "agent platform" and still was not in production. None of those months had been spent on their actual product. They had been spent on sandboxes, retry logic, credential handling, a custom evaluator, and an observability layer that mostly worked.
Then I watched a Code with Claude London workshop on Claude Managed Agents and realized almost every line that team wrote is now an API endpoint you call from any Anthropic key.
The mental model is four primitives.
Agent is the template. System prompt, skills, tools, MCP servers, per tool permissions. File read can autoexecute, bash needs explicit approval.
Environment is the sandbox template. Network access, preinstalled packages, which infrastructure runs it. Anthropic ships their own. You can also bring Cloudflare, Modal, Vercel, or your own fleet for data that must not leave your VPC.
Session is the ongoing conversation. Same idea as opening claude.ai, except you own it through the API. Preload GitHub repos or files at creation.
Event is the wire format. User events in, agent events out, session events for lifecycle, span events for long operations.
The feature worth the most attention is outcomes. You write a rubric. Claude builds the answer. Claude grades itself. Claude iterates. Claude grades itself again. It loops until it passes. Most teams build this by hand in two sprints. Outcomes makes it one event.
What you stop building. Your own agent loop. State recovery. Skill and MCP integration. Durable storage for sessions and events. A sandboxing fleet that reacts to Claude. End user authentication done securely. Credential vaults that inject MCP tokens without ever entering Claude's context window.
The honest version of "we built it for you" is that the parts that are easy to start and hard to finish stop being yours.
Use this when your product is the experience on top. Deal flow tools, internal copilots, customer agents, domain workflows. Skip it when your product is the platform itself. Frameworks, eval harnesses, raw research infrastructure.
The hard middle case is teams that have already built half a platform. Write down what is load bearing for your product, what is customization, what is rebuild. Map against what Managed Agents now does. Most teams find more rebuild rows than they want to admit.
The plumbing is done. The interesting work is what you do with it.
What is the part of your agent platform you would not give up?