AWS DevOps Agent Managed Agents: See What You've Learned (and why!)

You're working an incident, and the AWS DevOps Agent investigation in front of you cites a memory that looks out of date. Before you act on it, you want to know how your agents learned it. You want to know which agent wrote the memory, what makes that agent run again, and what else it can read. Today we're releasing a page for every managed agent , so you can answer those questions from the agent itself.

In AWS DevOps Agent, the author is usually the learning agent . It's the managed agent that writes and refreshes most of the knowledge in your Agent Space. Until now, you could see the knowledge the learning agent produced, but not the agent behind it. The learning agent's page shows its runs and its schedule. Every managed agent's page shows its instructions and the knowledge it reads.

This post explains the model behind the pages, then follows one memory through them. Nothing needs to be migrated, and your knowledge stays where it is. To find the pages, open Agents in the navigation and select the Managed tab.

An Agent Space  is the workspace where AWS DevOps Agent keeps your investigations, chats, and knowledge. That knowledge takes three forms. Instructions , kept in an AGENTS.md file, tell an agent how to work before it starts. Skills  are procedures an agent can load while it works. Memories  are what an agent remembers across runs, and they live in memory stores.

Two kinds of agents read that knowledge. Custom agents  are the ones you create. Managed agents are the ones we provide and maintain, such as the agents that triage incidents, find root causes, and answer questions in chat. You can write knowledge yourself, in custom skills and memory stores of your own. The learning agent maintains the managed memory stores  your agents read, refreshing them automatically from your cloud resources, application code, investigations, and chats. If you know this knowledge as learned skills , it now lives in these stores.

An Agent Space  is the workspace where AWS DevOps Agent keeps your investigations, chats, and knowledge. That knowledge takes three forms. Instructions , kept in an AGENTS.md file, tell an agent how to work before it starts. Skills  are procedures an agent can load while it works. Memories  are what an agent remembers across runs, and they live in memory stores.

The Knowledge page answers one question well: what knowledge exists. As an Agent Space learns more, a second kind of question becomes more useful: who produced a piece of knowledge, and what will change it. The three questions at the top of this post are that kind, and each one is about an agent.

We're the developers who build AWS DevOps Agent, and we tested that idea before building on it. We traced one incident through the people and agents involved and marked what each of them could find. Every path led to an agent. We also interviewed an on-call engineer from an Amazon team that uses AWS DevOps Agent every day about how they think of agents and knowledge. They expected to find the learning agent alongside the other agents. In their view, agents are the machinery, with triggers, history, and output, and knowledge is what runs on that machinery. They also expected the instructions every agent shares to sit with the agents.

That is the model the pages follow. Each agent has a page, and you reach the knowledge it reads from there. The Knowledge page is still the full inventory for when you want to see everything at once.

Say the memory you're chasing sits in the understanding-pipeline-topology store. Open Agents, select Managed, and choose the Learning agent. The Managed tab groups agents by what they do, into Learning, Chat, Incidents, Improvements, and Releases, with Learning at the top. Each page carries a Managed badge.

History  lists every run the learning agent has made in your Agent Space. You can search runs by title, filter them by status, or sort them by when they were created. When a run refreshed a memory store, its title links to that store. Open the run to see what the agent looked at and what it concluded.

Triggers  shows when the agent works. It lists the seven scheduled tasks that keep your managed knowledge  current. Each task names the memory store it refreshes, links to it, and shows when it last ran in your Agent Space. Select a task to see what starts it.

Find your memory's store in the table. It tells you the learning agent wrote it, that it refreshes based on activity in your Agent Space, and when that last happened. That settles the first two questions.

Knowledge  settles the third. It lists the skills and memory stores the agent can read, each marked Managed if we maintain it or Custom if you created it. Select an entry to open its detail page. The Knowledge page tells you which managed agents use a skill or memory store, and this tab tells you what one agent can read.

Agent Definition  holds the instructions the agent follows before it starts, in two sets side by side. The global instructions apply to every agent, and the agent's own instructions apply to it alone. You can edit both. Saving the global set asks you to confirm first, because the change reaches every agent session in your Agent Space. The agent's own set saves without a prompt, since only that agent is affected.

Managed agents are ours to maintain, and the pages reflect that. You can read the Triggers schedule but can't change it, and you can't start a managed agent from its page or change its tools. What you control is what the agent reads, through its instructions and the knowledge you add.

Only the learning agent has History and Triggers tabs. Every other managed agent shows Agent Definition and Knowledge. The Knowledge tab shows what an agent reads, and you can't add knowledge to the agent from it. The engineer we interviewed asked for that, and this release doesn't include it. The agents you see also depend on your Agent Space, since some belong to features you might not have turned on.

 

Back to the investigation and the memory that looked out of date. You can now name the task that refreshes it, see when it last ran, open the run behind it, and read the agent's instructions. With that in hand, you can decide how much weight to give the memory before you act.

To try it, open Agents, select Managed, choose the Learning agent, and start with Triggers. If a task surprises you, or a page leaves a question unanswered, we'd like to hear about it.

For the full reference, see Managed agents  in the AWS DevOps Agent User Guide. For how custom agents compare, see Working with AWS DevOps Agent custom agents  in the same guide.