Initially the thought was that single AI agent would handle arbitrarily large workflows. Instead, the pattern that seems to be working is deploying subagents that have specialization by task to avoid context rot. AI agent division of labor may be the future.

There’s a window right now where AI agents will get built for every vertical and domain. The playbook is to go deep on the context engineering required for the vertical or particular space, figure out the right UX that ties into the existing workflows naturally, and connect to the relevant data sources and tools. Especially early on, it’s useful to get as close to key customers as possible to figure out what’s working and what’s not and constantly make improvements to bring them back to the mothership. AI is moving so fast right now that there’s a huge premium in making quick updates and seeing how they improve the customer’s workflows. It’s also important to price the agents for maximum adoption with simple subscription prices or on clear consumption model, and expect to ride out the cost improvements from AI efficiency. Don’t get too greedy on price right now as market share is likely most important. It can be helpful to go after use cases that are constrained by the availability or high cost of talent. This means that any incremental boost in productivity in these spaces offers high ROI for the customer. In these areas, customers will always be willing to try AI agents to finally get around to solving their problems. This is why AI coding agents, security agents, or legal agents are taking off right now initially. These are all areas where demand for solving the problem has always exceeded the level of talent available. But every vertical has examples of this. There’s a clear moment right now where the next generation of these AI Agents will get built across every space.

If you imagine every job function and workflow will have an AI Agent counterpart to automate work, it’s clear we’re in the earliest stages of the AI wave. Every vertical and line of business represents a new market category and brand new TAM that didn’t exist before.

Division of labor for AI Agents will be critical for maximizing the impact of agents in all areas of knowledge work. We've long had a division of labor in organizations because it turns out having individual experts handing off tasks to each other is more effective than a bunch of generalists trying to do things a different way each time. AI Agents present the same dynamic. For AI Agents to work, you need just the right amount of context about the task that they're trying to complete. This means a deep domain understanding, set of knowledge to work off of, clear instructions, and set of tools to use. Too little context and the agent will fail. Yet, equally, as more of this information enters the context window, we know that the models can become suboptimal.  For a complex business process, if you put all of the documentation, description of the workflow, and instructions into the context window, we know that the agent will eventually get confused and deliver worse results. The logical architecture then in the future is to divide agents up in atomic units that map to the right types of tasks and then have these agents working together to complete their work. We're already seeing this play out effectively in coding agents. There are more and more examples emerging with people setting up subagents that all own specific parts of a codebase or service area. Each agent is responsible for a part of the code, and there is agent-friendly documentation for the code. Then as work is needed in that relevant area of the codebase, an orchestrator agent coordinates with these subagents.  We could see this pattern likely applying to almost any area of knowledge work in the future. This will allow AI Agents to be used for far more than task-specific use-cases and extend to powering entire workflows in the enterprise. Even as AI models improve to be able to handle larger context windows, and the intelligence levels go up, it’s not obvious that this architecture ever goes away. It’s likely that the role of each agent expands as capabilities improve, but clear lines of separation between subagents may always lead to better outcomes.

AI agent division of labor may map to human division of labor. People are using subagents that own specific micro services in a code base. Each service has agent-friendly docs that it updates as it goes. This way you optimize what’s in the context window. The future is wild.

The mental model for advanced AI is that the models will be highly generalizable, but they need to be paired with very specific context. Precise prompts, subagents divided up by task, agent-friendly docs, knowledge access. Even as models improve, this doesn’t change.

If you’re building AI Agents, it’s super important to figure out the optimal use-cases that maximize what agents are good vs. what they’re not ready for *yet*. There are so many categories of work that AI Agents can help automate or augment. Choosing the right ones that can deliver value in the near term and get better over time with model improvements is critical. Here are a few characteristics that seem to be working right now: * Work that requires a heavy amount of unstructured data and information. This could be documents, visual data on a screen, video content, and more. This is the domain that computers and software have never been able to do before, and the use-cases here are vast. * AI Agents are useful for things that otherwise require human judgment or interpretation, and that may always be the case. The moment you find yourself hoping to replicate something with very strict rules that happen over and over again, you probably want software, not agents. * The more complex work that’s being automated, the more that there’s a need for a human in the loop element. This is why code agents work super well right now is you can eventually test and study the output of the agent to figure out what came back right or wrong. Even when these agents do things wrong, intervention is relatively straightforward for any skilled user. * Bet on use-cases where the core intelligence of models getting better will continue to accrue to your agents. If you can solve everything about your use-case with AI today, it’s probably not an interesting enough market to go after. Go after scenarios where there’s incremental value that gets added with model improvements. Tons of more characteristics determine which use-cases are good for agents at this stage, but ultimately tons of opportunities in every category of work to go after.

One of the big upsides of AI Agents for knowledge work is the ROI changes dramatically on a number of things that you couldn’t have done before. There’s tons of work that we don’t do today because we can’t justify the “fixed cost” of getting it going. Almost every new idea becomes a meeting, with follow ups, and more coordination tax. So you, rightfully, prioritize only the highest impact work, and pray that you’ve made the right call on what that is. AI Agents changes the calculus here. The product team can afford to prototype more ideas to see which one is better. The business analyst can comb through more customer data to find a hidden insight. The engineer can build features faster. The legal team can better support smaller customers. The product marketer can run more campaigns or test more messages to reach more customers. Some of these things won’t matter a ton, of course. But many will. And by lowering the cost of trying a new idea, testing a marketing message, or researching a market, companies will start to do far more than before or at least get to their next destination faster.

The Figma case study is a great one. A team 100% focused on building great product, product led growth, fair pricing, building strong network effects, and expanding into logical adjacencies without getting too overextended. Many lessons for the next crop of startups.

Hard to overstate how slow the product design process was before Figma. You’d just brainstorm with low fi wireframes, feedback loops would take forever, and you’d wait weeks for mocks. Now you basically see a prototype of your product within hours of coming up with it.