Anthropic Managed Agents - building the block of flats

The organisations that are having the biggest success with managed agents, don't start by building the agent.

They start by building the block of flats that the agents will move into first - the architecture, the plumbing, the electrics...

Three things to focus on:

MCP Server

This is a separate project that connects your future agents safely and securely into your other systems.

Email, chat, CRM, support, ERP, internal files and folders.

This provides the security and governance around what your agents can see and do - it is very deterministic (meaning the same input will always get the same output).

Even if one of your future agents decided "Let me access all of our open invoices" if the MCP Server doesn't expose that tool to that agent, then there is no invoice information visible to the agent.

If an agent says "I'm going to email all of our customers" but the MCP Server only allows communications to email addresses with your company's email domain - no external emails are being sent.

We recommend an individual or team is responsible for this, and whilst that could be the same person building the agents, this is its own project with its own deployment and tests to maintain a clear firewall between what the agent wants to do and what the agent can actually see and do.

Communication channels

It's worth thinking about how your people (or customers) will be communicating with your agents.

Email? Teams/Slack? A web interface? Via an LLM like Claude or ChatGPT? An update in another system?

Picking the right channels allows you to put your agents close to where the work is happening.

We use Vercel's ChatSDK either directly or as a reference - it has adapters for Slack, Google Chat, Teams, Email, Whatsapp, Discord and others.

Giving agents a human persona (name, avatar, email, chat, folders) makes it much easier for your colleagues to understand and interact with them.

We set our agents up as users in Google Workspace - they get email, chat, calendar, Drive. We create an avatar, give them a name, and our human colleagues message Karolina, Brian, Krista, Deirdre in just the same way they would message a human.

We include our agents in our email groups - all@kowalah.com which keeps the information flowing between them and acts as triggers for other agents to do work.

(Karolina our project manager will pick up on new opportunities raised in a client meeting via Granola notes, which Brian our AE will pick up and turn into a new deal in HubSpot)

Folders and Memories

Giving your agents context allows them to be more effective, and to get better over time (just like a human employee).

Set them up with their own Google Drive or OneDrive where they can create their own folders and documents, read their own documents, and read from the folders of other humans and agents.

Set them up with multiple memory stores:

• Organisation: (read only) a memory store you manage directly that has all the context about your company, your strategy, your goals, your metrics - think of this like the intranet your human colleagues would go to to learn how the company works.
• Agent: (read/write) a memory store where each agent will build its own memories about the work it is doing, the world around it, and the humans and agents it interacts with.
• User: (read/write) a memory store used across all of your agents where each agent stores and reads information about your human users, what they are working on, and their preferences.

Once you've built the block of flats, then you can start to think about moving agents in.

You now have the structure that will allow your agents to flourish:

• They have safe, governed and independent access to your systems
• They can be contacted via the tools your team already use
• They have context that will continue to grow as they onboard

We're excited by what is to come with Managed Agents and will continue to share our learnings here.