I agents are running wild in most businesses right now. And if they aren't running wild in your organization yet, just give it a minute.

As a result, every day I get a ton of pitches from startups promising to be the solution for organizations to track, control, and manage AI agents within their company networks.

Earlier this month, Anthropic launched its version of a get-out-of-jail-free card with Claude Managed Agents, which "pairs an agent harness tuned for performance with production infrastructure."

But there were two drawbacks. One is that because it's in the cloud, you have to pay Claude's high token costs. That hasn't changed. You're still basically trading those token costs for the time and expense of building your own agent safety infrastructure.

The other drawback was that it didn't retain memory between sessions. That's no longer the case as Anthropic has launched built-in memory for Claude Managed Agents in beta. And there's a lot to like.

A few things to take note of:

• You control the memories: The memories are now stored in files that you can export and share, and you can manage them through the API, which makes them very enterprise-friendly
• Agents can share with each other: If you allow it, agents share the context they are learning with each other to make them more effective over time
• Teams have already been testing it: Netflix has been allowing agents to carry context across sessions and it's allowed teams to stop manually updating so many prompts; Ando (which is building a Slack competitor) is using the feature to retain organizational context for customers; and Rakuten is using the feature to reduce first-pass errors by 97% for its task-based long-running agents

Our Deeper View

If you're going to bring agents into your organization, you'd better have a solid plan for safety and security, because agents are non-deterministic systems that will do whatever it takes to accomplish a goal a human gives them. And that can include bypassing all kinds of policies, safeguards, best practices, and common sense. But once you have all the proper safeguards in place, you'll also want to let the agents cook and not have to re-explain the same context over and over again. That's what Anthropic is solving for with this new memory feature for Claude Managed Agents. Expect others to follow.