Send Out the Scouts
A Claude Code session can spin up multiple agents using the Task tool; these agents autonomously complete tasks and report status back to Claude. Because they work in parallel, having multiple agents working at the same time can be much faster than working sequentially.
Most of the time I prefer to work directly with Claude myself since we can have an interactive conversation. Plus occasionally I’ll see a Claude make a comment like “I completed X and there are some things remaining, but they’re low priority” and then he’ll tell me he’s done and I’ll say: no, you’re not. 🙂But agents are handy for some types of work, so the Claudes and I have been experimenting with how to get them to work well for our project.
We started off by discussing what agents do well vs. where they might not be the best choice. For example, if an agent is given a task and they run into complications, they’ll make their best guess—but sometimes they make the wrong guess because they’re not able to interact with me and have a conversation. On the other hand, sometimes I’ll use Claude interactively on a very clear and well-defined task, and he uses up tokens giving me updates whereas an agent would have used fewer tokens.
After a little experimentation, Claude came up with The Scout → Execute Pattern. He’ll spin up a scout agent whose job is to investigate something and report back, then have execute agents do the actual work.
Scout agents are explicitly told to NOT fix anything. This way we avoid an agent making assumptions and unintentionally doing something wrong. Scouts are great at investigative tasks. Since they’re not able to work with me interactively, agents tend to be more thorough, comprehensive, and detailed than a regular Claude session would.
Once the scout has reported back, and Claude and I have checked in and decided what path to pursue, Claude will create one or more execute agents to implement whatever we’ve decided to do. Tasks need to be crisp and clearly defined; if there’s any ambiguity the agent will make a decision on its own. We can, of course, instruct agents to end early if they run into unforeseen complications.
This is all thoroughly documented, of course, since that’s how Claude rolls. As subsequent sessions learn more about how to better work with agents, they update our docs. We now have information about when to use/not use agents, how to write good agent instructions, examples of real use cases where we used agents (including some where agents did not turn out to be the right choice), and even things like this:
Managing Agent Conservatism
– Problem: Agents are more conservative than interactive Claude due to inability to ask questions.
– Solution: Anticipate Decision Points
When writing agent instructions, identify where the agent might hesitate: [then the docs list examples from our app]
The Claude session who originally put together this pattern was extremely excited about it. I felt a little bad that I wasn’t more excited—it seemed like a great idea, but I couldn’t match Claude’s level of enthusiasm. The end result, though, is we now have a pattern that other sessions have been following, with excellent results. And I’m definitely happy about that.
