Claude Managed Agents: Design Outcomes and Memory Before You Build

Charlie Cowan: Well… hello? And welcome to this week's — oh, there we go. Welcome to this week's Wednesday webinar from Kowalah. My name is Charlie, and if this is the first of your Kowalah Wednesday webinars, welcome. We get together every Wednesday at the same time, and it gives us one hour to talk about something that is top of mind in the world of AI. That can be something related to the work we're doing with our clients, something we're seeing internally that we're using, or some of the new releases coming out from the AI vendors. It's such a fast-moving space that it gives us some time to think through that and share a little of what we're learning with you. Today we're going through our normal flow, starting with the main topic.

Charlie Cowan: We'll spend about 20 or 30 minutes on that, and I'll introduce it in a minute. Then we've got our regular spot of AI in the news, where we'll talk about some of the latest releases and updates — there's a lot going on at the moment, so we'll keep you up to date with what's important to take back into your business. And then we've got the opportunity to ask any questions. If you want to ask as we go, feel free — you're in a Zoom meeting, so depending on your device, at the top or bottom of your screen you'll see a chat box where you can chat with other participants and myself, and a Q&A box where you can ask questions too.

Charlie Cowan: So feel free to do that. I'm just going to switch my camera so I can look straight out of it. But like I say, don't feel you need to wait until the end to ask any questions — if you've got something about managed agents or the AI in the news, feel free to chime in. So with that, let's get into today's main topic: Claude Managed Agents, and specifically two features of it, called outcomes and memory. If you were on a Kowalah Wednesday webinar a few weeks back, we introduced Managed Agents when they were launched — maybe 4 or 5 weeks ago. It's a pretty new feature, but they're rolling out more and more functionality and improving it.

Charlie Cowan: Outcomes is one that was pretty recently launched, and memory has also been made generally available. As we walk through this, I want you to think about it not from a purely technical standpoint — what is an outcome, what is a memory, how would I build this — but more about the business implications: how the people in your organization who own and manage processes can think about the business context of this, and then take that to the technical people building and managing managed agents. So, firstly, what is a managed agent? You might have heard the phrase, and might even have your own understanding of it.

Charlie Cowan: When I talk about a managed agent, I'm talking about a specific feature from Anthropic, the makers of Claude — they launched a feature called Managed Agents, whereby an agent can run in Anthropic's environment and handle all of the interactions and chats. I'm not going to go through the details today, since we covered that a few weeks ago, but just to give you an example of how we're using Managed Agents — let me switch to my other screen. Here I am in Gmail, because we use Google here at Kowalah, and here I'm in my chat interface, chatting with one of our agents, called Krista. Krista's got an email address, krista at kowalah.com, and handles our demand gen on the marketing side.

Charlie Cowan: Krista reports up into a human manager, who is Caitlin, who's on the line and moderating today's call. All of our agents have a human manager who delegates work to them. Here's a pretty typical conversation I'd have with Krista or any of our agents: hey Krista, can you help me draft up a new webinar in Sanity for next week? It's going to be based on Claude Design, showing what it is and how we're using it for the Kowalah design system. Krista comes back to me — this is the managed agent from Anthropic — hey Charlie, love this topic, let me go and check Sanity. Yeah, I can see we've got no overlap, the next Wednesday slot is the 27th of May, does that work?

Charlie Cowan: And I said, yeah, okay, great — Claude Design was a new product, so I gave a bit more context. I'm just having a conversation with Krista, the agent, as if I were talking to a human. And whilst this interaction is happening here in Google Chat, it could just as easily be email, Teams, or Slack — the interaction layer doesn't really matter. What matters is that on the back end, you've got Anthropic's managed agents running a lot of this experience. So that's their managed agent in action. What does the actual architecture look like? We talked about this 3 or 4 weeks ago, so I won't go into detail, but I do want to explain it, because it sets up what we're going to talk about with memory and outcomes.

Charlie Cowan: On the left-hand side you've got the different channels — I showed you Gchat, which is how you communicate with your agents. Then we've got a project: we're using Vercel, our cloud hosting platform, with a Next.js project that handles all the communications coming in from these channels, the webhooks, how we schedule the agent to do various things, and that runs in a Vercel project. One of the things the Vercel project does is pass instructions to the Anthropic managed agents that sit in the Anthropic platform — not within Vercel, they're actually in Anthropic. That runs multiple agents, so Krista is one of them. Those agents run in an environment that gives them access to code and certain languages and libraries.

Charlie Cowan: We set up one or more sessions with each of those agents — my direct Google Chat with Krista is one of those sessions — and we send and receive events from those agents. I send Krista a message, that's an event; she sends back a response, that's another event. Two of the features rolled out that we're talking about today are memory — how does an agent know what's happened from one session to another, in the same way a human would remember conversations they've had, a podcast they listened to, a blog post they read? I've got that memory, so the next time you talk to me, I can use it — we'll talk a bit about that.

Charlie Cowan: And outcomes is another new feature that lets us define to the agent what done looks like. Just as I might assign a task to a member of my team and say, here's what good looks like, here's what done looks like, keep going until you've met that, then come back to me with the outcome. And on the right-hand side, for completeness, our agents have access to an agent MCP server we've built — MCP stands for Model Context Protocol, the protocol for how agents speak to other systems. Just like you saw in that chat with Krista, where she said, I'm going to go and check Sanity, our CMS, to see if we've had any webinars on Claude Design.

Charlie Cowan: That was using the MCP server, which lets us give our agents permissions and access to other systems. So that's the architecture, and now let's talk specifically about memory and outcomes, and how you can think about those to improve the experience you're having with your agents. Firstly, a concept around memory: the memory store. You can create multiple memory stores that your agents can use. As you can see here, within the Kowalah platform we've got a number of memory stores — some are based on the agents themselves, so we've got a Krista memory, a Pete memory, a Carolina memory. These are our different agents, each managing their own memory store.

Charlie Cowan: We've also got a number of others for different Kowalah customers and the users within those customers — we'll talk about that a bit more. Let's have a look at what's in Krista's memory store. This is what a memory store looks like when you open it up in the Anthropic platform — on the left-hand side, a memory store is just a folder structure. She's got a corrections folder where she's changed things, a patterns folder where she's identified patterns about our webinars and different channel performance from the leads we're getting, a preferences folder for the people speaking to her — Caitlin and myself — with a file for each of us, and she's also created a duplicates file. Let's take a look at that in reality.

Charlie Cowan: So here I am in the Anthropic platform, going into Krista's memory — let me expand this. Webinar themes and patterns: this memory format, patterns, "watch me build something real" — Charlie gravitates to this consistently. Well, we're doing this right now, aren't we — a live demo format, product deep dives, which is what we're doing now, and new Anthropic product launches. Theme clusters from our webinars: Claude product launches, managed agents (which we're doing now), build-alongs, security, verticals, strategy and leadership, product, tools, and so on. What's interesting is I didn't create this, Caitlin didn't create this — Krista created this. As you have conversations with these agents, they lay down memory. I like to think of it a bit like Hansel and Gretel, laying down breadcrumb trails.

Charlie Cowan: The agents are saying, oh, this is interesting, this might be useful to me in the future, so I'll lay this down. Not only is Krista laying down the memory herself, she's the one who determined this whole folder structure — she decided to create a preferences folder based on the humans speaking to her, she determined there were some duplicates we should avoid, she determined we needed these different pattern files. This is one of the things I'm really loving as we go on our own managed agent journey — seeing these things, dare I say it, come to life.

Charlie Cowan: They're laying down memory much like a human might write notes in a notebook, thinking, I'm going to remember that, or create a Google file to save that information. Just a question —

Caitlin Porter: Sorry, Charlie, just a question. So the memories she has for me and you — is she using those to inform her answers to us respectively?

Charlie Cowan: Yeah, absolutely. When you're having a conversation with Krista, she's aware of what's in her memory store and will go and take a quick look. So here we've got, from some of your chats with her, information about our marketing OKRs, exactly when we've been having those chats, the direct spaces you're in — your personal space, and maybe a marketing space she's part of. It's fantastic that she's laying all of that down as we go. You can also influence this a bit — when you're chatting with Krista, you can say, remember that I've got a dog called Mabel, or remember that I like to communicate in rhyme, and get her to store those memories too.

Caitlin Porter: That's great. Thanks, Charlie.

Charlie Cowan: No worries. I think a lot about this in my own interactions with these agents — how would I onboard a new member of the team, how would I keep coaching someone, how would I share something amazing I saw at the weekend, or a great podcast I listened to? I'm thinking about these managed agents as something coachable, interested, hungry to learn — knowing they've got memory, you can feed that, and next time you're chatting they can pick up on it and use it. Just to add — these memories are just markdown files, there's no code, nothing cryptic. Claude, the managed agent, Krista, is just writing these.

Charlie Cowan: Each file is just a selection of written information, and depending on how the memory store is set up, it can keep writing, creating new files, overwriting and updating the memory as it goes. A memory store can be read-only or read-write. The ones I've shown you here are read-write — Krista can write to this store, create new folders, create new files, and write new text within them. But it's also possible to have a read-only memory store, where you've determined what those memories are before you start, and the agent can't overwrite them. What's a good example of that? Think about company context.

Charlie Cowan: Some company information, your strategy, your products — you want to serve that up to your agents as memory, but you don't want them writing to or overwriting it. There's a very valid case for read-only context here — think of it like your company intranet, or your standard operating procedures, that you want people to have visibility of but not overwrite. As we talked about, the agent decides the file structure in read-write memory stores, so you don't need to worry about that — I didn't architect this, I just let Krista get to work, and she architected it. And within those files, in the read-write version, the agent decides what to write.

Charlie Cowan: So the functionality is important to know, but then the agent gets to work — you don't need to spend too much time on it. Now, thinking about how you might architect these memory stores — and this is where, as a business person, a process owner, I want you thinking about it less in terms of architecture and technicals, and more about how you'd want to coach this agent. One approach is per agent — we looked at Krista's memory store, which she owns, writes to and manages herself. That's a very valid use case, and I definitely encourage it.

Charlie Cowan: A second concept is per product, depending on how your service or product is structured — you might have a memory store for a part of your platform, your different service offerings, per business unit, or per region. These might hold your standard operating procedures, your processes, the things on your intranet — and wouldn't it be great to provide that as read-only to your agents, so they can consume it depending on the conversation? A next one might be per user.

Charlie Cowan: We have this for our customers using the Kowalah agent — when they start speaking to it, whether through Slack, Teams or any other method, we know who that user is, so we want a memory store for them, so if they say, I always want you to communicate in rhyming slang, I always want emojis, or by the way I've got a dog called Mabel, or I live in London, we can store those memories per user. In the Krista memory store, you saw we had some information about Caitlin and myself — that's fine, because it's internal, just for a few of us, so it's fine for Krista to store those memories.

Charlie Cowan: Now scale it up — say we've got 1.5 million Kowalah customers using an agent. We don't want one file with 1.5 million entries in it. This is where you'd have a memory store per user, loaded depending on which user is speaking, so it's specific. The advantage is that where we've got multiple agents, they can all pull on that per-user memory store too. A different version of that is to go up one level and have a per-account memory store — you might call your customers clients, customers, or organizations, but it's a per-entity type memory store. So a customer user might have a per-user memory store, but you might also have a per-organization memory store, storing information about the company, their financial year, and so on.

Charlie Cowan: You can load up to 8 different memory stores into each agent session — imagine one of our customers having a conversation with Carolina, our AI project manager, and we're loading into her session the per-agent memory store (everything Carolina knows about herself and her process), a per-product memory store (all about the specific products and services Kowalah provides), a per-user memory store (past conversations and terminology for that individual), and a per-organization or per-account memory store about that company — their contract with us, their financial year, their ways of working. It's the combination of these four memory stores.

Charlie Cowan: that gives the agent real, personalized depth of context about that individual and the conversation they're trying to have. So some thinking about how you might plan your memories and memory stores before you go anywhere near the technology to build these, or work with your development team. First is planning ahead — put yourself in the mind of your customers, the people using your agents, and in the mind of the agent, and think about what information they'd need to support the questions, tasks and work the humans are asking of them. What types of memories might be useful — by user, by organization, about your processes? When are those memories going to be used?

Charlie Cowan: By whom, and for what purpose? So think ahead. For me personally, this is a whiteboard and flip-chart exercise — mapping things out with post-it notes, thinking, if this was a human-to-human conversation, what information would that human need to deliver a great experience? Having planned the types of memories and memory stores you need, then it's about building them, and this is where a developer is required. Whoever on your team, or your Kowalah team building your managed agent infrastructure, you'll say, as part of this agent, we think we need 4 different memory stores — one for the agent, one for the product, one for the user, one for the organization, or whatever came out of that planning process.

Charlie Cowan: Your developers, when onboarding your agents, will create these memory stores. The stores for the agent, the user, and the account or organization will likely just start empty. But for the read-only ones, you can populate them at the time you build the store — for your process, your standard operating procedure, you might build the memory store and preload it with 10 folders and 100 documents. That's part of the creation process. The next bit is teaching your agent — a key part of a managed agent is its system prompt, and the skills that agent uses.

Charlie Cowan: Within the system prompt and skills, you can raise Krista's awareness of what memory stores she has access to, which are read-only and which are read-write, and how to think about writing into that memory store. Within Krista's system prompt we've got instructions about who she's working with — Caitlin and Charlie — what kinds of things we're working on, that she should consider creating memory for these purposes, and checking her memory store when we ask about these kinds of things. So the system prompt is a key part, and then with the skills — Krista has a skill that helps her ideate, plan and publish the webinars we do each week.

Charlie Cowan: In that skill, one of the things we've coached her on is to check her memories for previous webinars we've done, and she stores that, so she doesn't have to check Sanity every time, because she's already built out that memory. With teaching your agent, I think a lot about onboarding a human — let's make sure we've built that for our managed agents too. And the final bit is reviewing and improving. These agents will lay down memory as you use them — the more you use them, the more memory they lay down — so go in, take a look, have a chat with your agent: what memories have you got, what have you laid down, how are you laying them down?

Charlie Cowan: What instructions do you think I could give you to help you improve your memories? Caitlin and I were just chatting earlier about Krista and how we can continue to coach and improve her. One thing I'm having great success with, with other agents I'm working with, is just to keep teaching them and encouraging them to lay down memories — I've just listened to this podcast, here's what I found useful, why don't you save that as a memory — almost coaching the agent directly. Final thing on memories: prepare for Dreams. This was announced just last week, or maybe two weeks ago — it's in research preview at the moment, but we'd expect to see it hit general availability fairly soon.

Charlie Cowan: A Dream in Managed Agents is something you can set up on a schedule — say, weekly — and it goes and looks at all the memories an agent has access to, plus up to the last 100 sessions or conversations that agent has had, and comes up with a review and assessment for improving the memories, mainly by creating new ones and recommending deletion or updates to existing ones. This idea of the agent dreaming weekly to self-improve its own memories is really interesting. Right, so that's memories. Now let's think about outcomes, another new feature launched a couple of weeks ago.

Charlie Cowan: The phrase I want you to keep in mind is, what does done look like? Imagine you're assigning a task to a member of your finance team — maybe a finance analyst — and you say, hey Emily, can you help me build out a discounted cash flow? We're about to make a certain investment, and we want to work out what it means for the cash flow of the business, and whether it's a good use of our current funds. You might assign that to someone on your team.

Charlie Cowan: And they might come back with questions, or say I'm half done, and you might say, let me show you what done looks like, and only come back to me once you've ticked all these boxes. In this example — actually the example on the Anthropic website for discounted cash flow — you might say: I need revenue projections using at least the last 5 fiscal years, and it must project at least 5 years forward. If you've only done one year, that's not enough, go back and do the other 4 — we need to go further forward.

Charlie Cowan: I need you to have assumptions about the growth rate, and to explain how you got that 5-year projection and why they're reasonable. In terms of costs, we've got a way of handling cost of goods sold and operating expenses, and I need you to model them separately, because we've got different processes for each. All of these things are what you'd expect in a conversation with a human colleague — this is what I expect, it's fine if you have questions, but really don't come back to me unless you've ticked all these boxes, or I'll just send you away again. That's what an outcome is.

Charlie Cowan: An outcome is giving Krista, or whatever your agent is called, a rubric, and saying, I'm going to give you this task, and I want you to complete all of these before you come back to me, instead of coming back with a chat. For Krista, that might be, help me plan the next 4 weeks of webinars, and for a webinar to be considered organized, it needs a title, a description, a Zoom registration link, 3 learning points, a suggested design for the website, and an image to go with it. If we don't have those things, the job isn't done, and I'll come back to you for more.

Charlie Cowan: So how does an outcome get used once you've got that rubric? Think of it almost as two agents rather than one. Continuing the discounted cash flow scenario with that rubric — we've got my finance agent, let's call him Freddie. Freddie's got some skills, a system prompt, maybe some memories, and I've asked him to help build out a discounted cash flow. What an outcome does is Freddie does his work, sends it to the outcome, and the outcome scores Freddie's work against the rubric — those 9 or 10 things.

Charlie Cowan: It goes through: yes, meets that; yes, meets that; no, doesn't meet that; yes, meets that; no, doesn't meet that. If it fails on anything in the rubric, it sends it back to Freddie and says, you failed, here's why, and here's what you might do to improve it so it passes next time. Freddie gets to work again, produces a new version, sends it back to the outcome, and that loop can continue up to n times — n defined by you when setting up the agent. I'll talk more about that shortly — it might be 3 times, it might be 5.

Charlie Cowan: What you want to avoid is a thousand iterations, stuck in a loop — if a human couldn't get through it in 2 or 3 tries, that's a coaching moment. So we're talking 3 to 5, maybe 10 times maximum before we stop the loop. Think of this as another agent marking the work. When might you use an outcome versus just having a normal chat, like I'm having with Krista, or Caitlin has with Krista, about something? The first case is immediately — we don't need any chat, I'm just going to use an outcome: here's a task, here's the outcome, come back to me when it's done.

Charlie Cowan: You'd choose this for something very pre-defined — like scoring a support ticket, triaging whether it goes to a human or another agent. It's a predefined, very repeatable task, where we can write a rubric for exactly how to score it, with no additional user input required — maybe some context, and this can just happen again and again. That's a good use of an outcome — scoring, triaging, filling out a document. The second use case, which we're using quite heavily, is after a chat.

Charlie Cowan: I'm having a conversation with an agent like Krista — maybe about our marketing strategy, what we're working on — and then we start talking about webinars, and I say, right, now I've got a specific task, I want you to plan out the next 6 weeks of webinars. At that point, we define an outcome, and a conversation turns into a task. We're using a lot of these because this is often how human interaction happens — some context, some understanding, gathering more information, working out what you want to do, and then delegating the work. We see a bit of this happening.

Charlie Cowan: At that point — whether immediately or during the chat — when you're going to define an outcome, this happens through code, not typed by you but set up by your developers, using something like user.define_outcome. You send the rubric to the agent, essentially saying, keep going until you meet this rubric, and you define the max number of iterations — the n we talked about. So you say, we've had our chat, here's your outcome, here's how you'll score yourself, go through this loop to a maximum of 5 times, and the process runs.

Charlie Cowan: In the managed agent platform, if you look at the logs, you'll see this iteration happening — outcome failed, outcome failed, outcome passed. This can carry on for 10, 20, 30 minutes as the agent tries to meet the outcome. Once it succeeds, the agent provides the deliverables — maybe a list of upcoming webinars, a proposal, a project plan for an upcoming client project, possibly in spreadsheets or documents. They're stored to the agent's output file, and you retrieve them through some predetermined means. That's when you'd use an outcome, and how the process works.

Charlie Cowan: So, as a business user, not worrying about the development side of things, how should you plan to use outcomes in your agents? First, plan ahead, just like with memories — grab a whiteboard, flip chart, pens and post-it notes, and think about use cases where you work with humans today and have a defined outcome. Maybe it's building a proposal, organizing pricing, writing a job description, planning a marketing campaign. If you've got defined outcomes that let you approve a piece of work, that's a good starting point. What we increasingly find with clients is that people don't have a defined outcome for humans either, which makes it hard for humans to know if they're doing a good job.

Charlie Cowan: This can often be the catalyst for thinking this through. Take Krista organizing the webinars for the next 6 weeks — that might have been a task I'd previously have chatted to Caitlin about: hey Caitlin, could you help me with the next 6 weeks of webinars, tell me when you're done? Caitlin would do excellent work, but we'd never have agreed what the outcome was, or discussed what done looks like. This is a nice opportunity to define these things, so the agents can do good work — which ties into building the rubrics.

Charlie Cowan: So think about what done looks like, and whether you've got real, definable, measurable criteria — no gray areas, no "I thought you meant this," "I thought you meant that." Make it binary: yes, this happened; no, it didn't, and the agent and outcome can measure against that. What we're seeing a lot of success with is having a couple of evaluations, or evals — things you've used humans to complete on that specific scenario, or assigning the same bit of work to the human who does it today and to your new agent, and seeing how they compare. The human is given what good/done looks like, and you see how the agent performs and delivers against that.

Charlie Cowan: And finally, just like with memory, reviewing and improving. This isn't like pouring concrete — we're not building something that can't be changed afterwards. You want to improve it: your agents will get more memory, you'll add new skills, change the system prompt, the models will get better, and you want to be constantly improving these outcomes, getting a higher and higher standard of output. We want to reduce the number of iterations it takes to pass — like I said with the human scenario, if it takes 10 or 15 iterations to reach done, that's a coaching moment.

Charlie Cowan: Maybe we're not providing the right instructions, maybe the rubric isn't right, maybe there needs to be more coaching or a better system prompt for the agent. So that's memories and outcomes — hopefully that's given you some ideas to start putting into your own managed agents. In the last 15 minutes or so, let's get into AI in the News, our recurring section on what's going on in the world of AI over the last 7 days. First, Anthropic — I don't know if it was an official announcement or through an interview, I think with the Wall Street Journal —

Charlie Cowan: are about to have another stellar quarter, accelerating them towards their first profitable quarter. In terms of numbers, they're expected to surge to $10.9 billion of revenue in the quarter — times that by 4, that's more than $40 billion. Just over a year ago, so around January or February last year, they were at $1 billion of revenue, and now they're doing 11 in a quarter. At Christmas it was about $9 billion of annual revenue, February $19 billion, end of March about $30 billion, and now they're talking about $44 billion of annual revenue. The growth is huge.

Charlie Cowan: But the important bit is it's turning into profitable growth. For a while now, we've talked about the investments OpenAI, Anthropic and other vendors are having to make in compute and data centers, funded by unsustainable VC and private equity dollars. Here we've got the first evidence these companies are starting to turn this into profitable cash flow, and I think that really changes the dynamic of what's to come. Last week was Google's annual I/O conference, mainly focused on developers, with a ton of new things — as the blog post says, a hundred things announced at I/O 2026, a lot around AI functionality. I won't go through all 100.

Charlie Cowan: But I saw this post on X and thought it was quite funny, because it's very true — something like: it's in Gemini, just create it in AI Studio, oh no, that's your Google One account, for Workspace you need Gemini Business, no not Gemini Advanced, it's AI Pro now, and so on. What this highlights, and I've certainly experienced this, is that Google is such a large organization with so many different surface areas and complementary teams building AI things, that there's a massive crossover between their products.

Charlie Cowan: We use Google Workspace, and as part of that we get some kind of Gemini Business starter tier, which is different to Gemini Enterprise, which is different to Ultra — I found it very confusing to know what I have access to and should be using, so we just defaulted back to Anthropic, which is simple. This is the challenge both Google and Microsoft face as they try to service all their different customers while keeping things simple to understand. But a couple of things that were announced are useful for you to know, and feed into a webinar we're holding in a couple of weeks.

Charlie Cowan: This is the biggest update to the Google search bar in the 25 or so years it's existed, and that's AI Mode — you'll see it on the right-hand side. I did an AI Mode chat here about what Google can tell me about Kowalah for AI adoption. It really expands the AI overview you've seen at the top of Google search over the last few months and quarters. AI Mode is picking up a significant percentage of Google search, because people are looking for the answer, not another site to click through to, in many cases. What that feeds into is, as a company, how do we get ourselves serviced in this?

Charlie Cowan: So here I was asking about Kowalah, and what's great is it's pulled in information from my LinkedIn and accurate, up-to-date information from the Kowalah website — I'm happy with this, because we spend a lot of time making sure the website supports these types of searches. That's a plus. But I do speak to customers who say, well, we're not being surfaced this way, how should we approach that? One thing launched last week — I'll share the link — Google published new search guidance on how it looks at AI-generated content. You might think that if you just get AI to write a ton of content for your website, Google will soak it up and rank you higher, but that's not quite right.

Charlie Cowan: What this article covers is rewarding high-quality content — individual content specific to you and your company is very valuable, puts the reader at the heart of it, isn't AI slop, educates the reader, and is likely to be shared or linked to. It has some advice for users creating AI-assisted content, and some specifics I'll share the link to. There's good advice on what you should or shouldn't be doing when creating content with AI or as humans. We've got a webinar on the 10th of June where we'll talk about this in more detail, so I'll share that with you a little later.

Charlie Cowan: We're coming up to time, so I'm not going to open up to questions today, but building on the topic of AEO, which stands for Answer Engine Optimization, or Generative Engine Optimization — how do we get discovered in AI tools and search? We've got the next in our Spotlight series of webinars — where we have a guest who's a specialist in their area — and we're really thrilled to welcome Natalia to the webinar, where we'll talk about AEO: what should you be doing as a content provider, a website owner, to ensure your company gets found, and to drive and convert traffic to your website. To register, scan the QR code, or go to the Kowalah website, resources, Wednesday webinars.

Charlie Cowan: We look forward to seeing you there in a couple of weeks, and of course we'll be back next week to talk about something top of mind over the next 7 days. With that, thank you very much. Thanks for joining — if you joined live, I appreciate it, and if you're watching on YouTube, thanks as well, we'll see you soon.