Would be really cool if there was a noise cancelation / transparency setting on AirPods that could filter in/out specific kinds of sounds. Like I only want to hear nature, but not other people or cars…

CEO: How's that AI agent we built for the marketing team working out? Head of AI: Pretty well. It's saving Sarah about 10 hours a week. CEO: Amazing. What's she doing with all that extra time? Head of AI: Mostly maintaining the agent. CEO: Sorry? Head of AI: It breaks sometimes. Workflows change, APIs stop working, that kind of thing. CEO: How much time is she spending on that? Head of AI: Around five hours a week, let's say. CEO: Okay, so we're still saving five hours. That's good. Head of AI: Yeah, until we build the next agent. CEO: We should probably automate the maintenance then. Head of AI: We could. CEO: Great. Let's do that. Head of AI: And who's going to maintain those agents? CEO: ... Head of AI: Because Sarah's getting pretty busy.

Marketers spend enormous amounts of time inventing customers. Sarah is 35, has two kids, lives in the suburbs, and cares about convenience. Cool. But what does Sarah actually care about right now? At Waldo, we're experimenting with something called proxy audiences: identifying large groups of real people who represent or influence your customers, then tracking what they publicly discuss. Take Tin Can, an old-school landline for kids. We recently got one, and my five-year-old is obsessed. If I were marketing it, I'd follow parents of young kids to understand what they're talking about, which products interest them, and how their interests evolve. That's a much more current picture of your audience than a fictional profile you update every few months. I think marketers will increasingly build strategies around what real people are actually discussing.

All these personal AI assistants have a serious problem.. they're creating more work for the people they're supposed to save time. I've built a bunch of automations to help me run Waldo, and I've also experimented with things like Grokbot (which I was excited enough about to upgrade to a paid tier!). A few weeks later, I barely touch it. The recurring problem I've noticed is how much ongoing effort these tools require to stay useful. I had an automation tracking usage across our design partners. That was useful initially. But the information became less relevant over time as statuses changed, and the automation just kept on sending me useless updates. Each individual adjustment seems minor, but the effort adds up when you're managing several of these systems. I suspect this is going to create a massive retention problem for personal AI assistants. People love the idea of having an assistant that takes work off their plate. They're going to lose interest pretty quickly if keeping that assistant useful becomes another responsibility on their plate.

Synthetic personas (LLM focus groups) are overhyped. In tests we ran over the past few weeks, plain LLMs outperformed persona-instructed ones at predicting which marketing messaging would resonate. My working theory: a general LLM is already trained on a massive distribution of human behavior. Force it into one narrow persona and you pull it away from that signal. Tell me I’m crazy.

It’s amazing to me how many of the top AI companies aren’t AI-pilled at all internally.

The annual marketing plan is dead. AI is making everything move too quickly for a plan written once a year or quarter to stay useful for very long. Teams can launch faster, test faster, respond faster, and change direction faster. By the time you finish executing against one set of assumptions, the market may already look different. At our team offsite last week, we tried building around a different model. We created a system that starts with a company’s strategy and then keeps watching what happens around it. When something meaningful changes, the system looks at that change through the lens of the strategy and gives you a few possible ways to respond. Felt like Oregon Trail for marketing!

We got the Waldo team together in Niagara(!) last week for our latest company offsite. Most of us spend the year working together through Slack, Huddles, Meets, and a lot of tabs, so it was great to have everyone in the same place for a few days. We toured some vineyards, got our steps in at Niagara Falls, got completely soaked on the boat, ate well, and had enough time together outside of work to remember why we like each other. We also carved out time for an internal hackathon before everyone headed home. The team built a handful of experiments around problems we have been thinking about at Waldo, including a new way of approaching always-on marketing strategy and some testing around synthetic personas. A very good few days with a very good group of people. Already looking forward to the next one.

This is a post about AI assitants.

We’ve been building a lot more agents lately, and I’ve landed on a pretty straightforward process for getting from idea to working agent without overbuilding the infrastructure. If I were building an internal agent today, I’d do it like this: 1. Start in a managed environment. For us, that has been Claude Managed Agents. The goal at this stage is speed. Get the workflow working, make sure the agent is actually useful, and avoid spending too much time on infrastructure before you know the thing is worth keeping. 2. Keep the agent definition in your own Git repo. Prompts, manifests, configuration, and setup should live somewhere you control. That gives you version history, makes changes easier to track, and means the core logic does not live entirely inside the platform you happened to start with. 3. Push it into the runtime through the CLI. We use CLI tools to publish the version in Git into the managed agent environment. Git stays the source of truth. The managed platform is where the agent runs. 4. Port it later if you need to. Once the workflow is proven, you can move it to a more model-agnostic or cost-efficient setup without rebuilding everything from scratch. I would not spend weeks building flexible infrastructure on day one. But I also would not build an agent in a way that makes switching providers painful later. The process I like right now is pretty simple: move fast on the first version, but keep ownership of the part you may eventually want to move.

Our @linear has become a wasteland of impossible to read cards now that Claude is basically doing it all, and it's made linear unusable for our humans. I'd love if @karrisaarinen could share what they put in their claude.md or equivalent to work in @linear effectively as a team of agents repping humans. As of today I put in place a ton of structure in our repo's claude.md around when to use projects, how to title them, how to set priority, when to move stuff, etc. But would be great if linear had some guides on this...

We’re building marketing agents for a logo you certainly know, that’s running ~1,100 ads at a time. There's just no human way to track how different ICPs are responding across all of them. ... but agents can. Paid teams have spent years building systems around spend, conversion, creative performance, attribution, and audience performance. Meanwhile, a huge amount of useful audience intelligence still lives inside comments, replies, objections, questions, and the language people use when reacting to a message. Comments, replies, and other audience reactions will increasingly become inputs into marketing systems that teams analyze continuously instead of information someone checks manually when they have time.

If you want to find a good AI agent use case inside your company, spend time studying how one person actually does their job. We’re currently building managed agents for marketing teams, and most of the work starts with understanding the existing workflow in detail. To quote Einstein, “if I had an hour to solve a problem, I'd spend 55 minutes thinking about the problem and only 5 minutes on the solution.” One person on a team we’re working with manages millions of dollars in media spend herself. A role like that contains dozens of recurring processes, and each one is a potential place for an agent to help. 1) I’d start by mapping the work that happens every day or every week. 2) Then, I’d trace where that person repeatedly moves information between systems, where the same analysis or follow-up happens over and over, and where work gets delayed because one person has too much to keep track of. 3) After that, check whether an agent can actually access the data and systems required to do the job. We’ve already seen useful workflows become impractical because of permissions, security restrictions, or unavailable connectors. Once you understand the workflow at that level, you can define what the agent needs to watch, when it should act, what information it needs, and what a successful output should look like. The most useful agent ideas I’ve seen have come from studying real operational pain closely enough to find a specific job worth handing off.

Marketing in 2026:

Nobody makes more confident mistakes than an agent with the wrong context.

Building marketing agents 101: Your first marketing agent should own one recurring responsibility. Here is the process I would use to build it. 1. Choose a question your team answers repeatedly. For example: What changed in our competitors’ advertising this week, and what should we learn from it? A recurring question gives the agent a clear job and gives your team a way to evaluate its work. 2. Define the information it needs. For competitor monitoring, this could include active ads, previous campaigns, landing pages, audience comments, and your own performance data. 3. Decide what the agent should remember. The agent should know which competitors matter, what each one has done previously, which patterns have already been reported, and how your team responded. 4. Set the cadence. Decide how often the agent should check for changes and when it should report them. Some responsibilities require daily monitoring. Others may only need a weekly update. 5. Specify the output. Define what the team should receive, where it should appear, and who needs to see it. A useful competitor update might include: - The three most important changes - Why each change matters - Supporting examples - Recommended next steps 6. Establish the conditions for human involvement. Tell the agent which decisions require approval, which findings deserve an immediate alert, and when it should ask for more context. 7. Measure reliability before expanding its role. Track whether the agent finds meaningful changes, avoids repeating old information, and produces work the team uses. Once it can answer one recurring question reliably, give it another responsibility.

“AI can’t be creative” is going to age badly. But “AI should automate creative” may age even worse. AI is already capable of producing every component of an ad: copy, images, video, concepts, variations, and full campaigns are all moving in that direction. But using it primarily to increase output will create a lot more average work. AI makes production easier, but the audience still has limited attention. Instead, you should be using AI to create more entropy earlier in the process. An agent can continuously study customer conversations, competitor campaigns, cultural shifts, and performance data. It can identify patterns a creative team may never have had the time or information to find. That gives talented editors stronger raw material. - A strategist can develop a sharper brief. - A creative can explore an unexpected tension. The strongest use of agents may be surfacing the idea the team did not know to look for... a customer tension, competitor pattern, or cultural shift. It’s not unlike how AI is being applied to mathematical research. Sometimes you need to see move 37 to advance the field.