Session Slides

Welcome to Claude Dispatch. You have built agents. Maybe several of them. Now you're facing a new problem: requests come in from all directions, and you don't have a human in the middle anymore routing them. That's where dispatch comes in. A dispatcher doesn't build things. It reads what's coming in, figures out what type of work it is, and sends it to the right specialist. Today we're learning how to build that system so your agents can coordinate without you managing everything.

Here's the pattern. Most people build one big agent and ask it to handle sales questions, support tickets, compliance reviews, and marketing copy. That agent is slow. It hedges. It second-guesses itself because it doesn't have deep context for any single domain. Compare that to a specialist agent that only handles sales questions. That agent knows the product inside and out. It knows what questions lead to closes. It's faster and better. The reason is simple: specialization is more powerful than generalization. A dispatcher lets you build that specialized team without managing the handoffs yourself.

A dispatcher is simple. A request comes in. The dispatcher reads it - that takes seconds - and asks one question: what type of work is this? Then it knows which specialized agent to send it to. The dispatcher is not doing the work. It's just directing traffic. And because it's directing traffic to specialists instead of spreading the work across one general agent, everything gets better. Speed goes up. Quality goes up. The only thing the dispatcher needs to know is how to categorize the request and which agent handles which category.

Start here. Don't overthink it. Look at the last month of requests you handled. What patterns do you see? Most organizations have 4 to 6 distinct types of work. A sales team might have: lead qualification, objection handling, proposal generation, and deal structure. A support org might have: technical troubleshooting, billing questions, feature requests, and complaints. Write them down. For each one, ask yourself: would I train someone differently to handle this category versus that one? If the answer is yes, you've found a category. Your dispatcher needs to recognize these patterns. Your specialist agents need to own them.

Your dispatch map is a decision tree that an agent can follow. For each category of work, you write down three things. First: the trigger. This is what signals that a request belongs in this category. It might be a keyword. It might be a question pattern. Second: which specialist agent handles it. That specialist has been trained or configured to do exactly this work and nothing else. Third: what success looks like. You are defining the output quality that the specialist should aim for. The dispatcher uses this map every time a request comes in. Read the request. Find the trigger that matches. Route to the specialist. Done.

Five years ago, you had a person reading every request and routing it to the right team. That worked. Quality was high because humans understand nuance. Cost was also high. Now your volume has tripled. That human is drowning. And here's the thing: that human was following rules, even if they didn't write them down. An AI dispatcher can learn those rules and apply them consistently at any scale. You don't lose quality. You gain speed and free up the human to do something more valuable than routing requests.

Don't try to build the perfect dispatch system on day one. Start with one category of work that you handle regularly and that would benefit from specialist expertise. Write down the trigger as clearly as you can. Maybe it's keywords. Maybe it's a pattern in how the request is written. Build a specialist agent that handles that work. Test it 10 times. Then add another category. Each time you add a category, you're making your system smarter and faster. And here's the key: your dispatch map is not set in stone. As your business changes, your categories change too. Update them. Keep the map alive.

This is where your AI program stops being a tool and starts being a system. You've moved from "I have an agent that helps me" to "I have a team of specialist agents that coordinate automatically." That's a fundamentally different level of capability. One agent answering questions is nice. A team of specialists routing work to each other and producing expert-level output at scale - that's when you start seeing real business impact. Your dispatch architecture makes that possible.