Here’s something that would make any manager laugh: teaching artificial intelligence how to delegate effectively turns out to be just as tricky as teaching humans the same skill. Maybe even trickier.

Anthropic shared fascinating insights from building their multi-agent research system in their recent engineering blog post, and buried in all the technical details is a profound truth about management that applies whether you’re coordinating humans or AI agents. The challenge isn’t just getting the work done: it’s getting multiple independent entities to work together without stepping on each other’s toes, duplicating effort, or wandering off in completely wrong directions.

When Anthropic’s engineers first started building their system, they ran into the same problems you’ve probably seen in any office. Early versions of their lead AI agent would spawn 50 subagents for simple queries (the equivalent of assigning your entire department to find a single phone number), or give instructions so vague that different agents would end up doing identical work while leaving critical gaps uncovered.

One particularly telling example they shared: when the lead agent was told to research “the semiconductor shortage,” different subagents interpreted this completely differently. One dove into the 2021 automotive chip crisis, while two others duplicated work investigating current 2025 supply chains. Sound familiar? It’s the same thing that happens when you tell your team to “look into our customer satisfaction issues” without being more specific about what each person should focus on.

The breakthrough came when they realized that effective delegation (whether to humans or AI) requires much more than just dividing up the work. Each subagent needed four critical elements:

1. a clear objective
2. a specific output format
3. guidance on which tools and sources to use
4. explicit boundaries around their task scope

This mirrors what great managers have always known. You can’t just say “handle the Johnson account.” You need to specify whether someone should focus on the contract renewal, address the recent service issues, or explore expansion opportunities. The clearer the boundaries, the better the results.

What makes AI delegation particularly interesting is how it reveals patterns we might miss in human management. Anthropic discovered their agents needed explicit scaling rules embedded right in their instructions. Simple fact-finding required just one agent with 3-10 tool calls. Direct comparisons might need 2-4 subagents with 10-15 calls each. Complex research projects required more than 10 subagents with clearly divided responsibilities.

Think about that for a moment. They had to teach their AI system something most managers learn through years of trial and error: how to match the complexity of your response to the complexity of the problem. No need to bring in five people for a task one person can handle in an hour, but also don’t expect one person to tackle something that really needs a team.

The tool selection challenge proved equally revealing. Just as you wouldn’t send someone to research industry trends using only internal documents, AI agents needed explicit guidance about when to use different tools. An agent searching the web for context that only exists in Slack is doomed from the start, just like a team member looking for sales data in the marketing folder.

Anthropic solved this by giving their agents explicit heuristics: examine all available tools first, match tool usage to user intent, search the web for broad external exploration, and prefer specialized tools over generic ones. In human terms, this translates to making sure everyone knows what resources are available and when to use each one.

Perhaps the most fascinating discovery was how they taught their lead agent to think strategically about research approaches. They embedded the principle of starting wide, then narrowing down. Exactly how expert researchers work. Agents learned to begin with short, broad queries to understand the landscape before drilling into specifics.

This maps directly to how effective managers approach complex projects. You start with the big picture, get a sense of what you’re dealing with, then dive deeper into the areas that matter most. The agents that jumped straight into overly specific searches often came up empty-handed, just like team members who get lost in details before understanding the broader context.

The parallel processing breakthrough came when they realized complex research naturally involves exploring multiple sources simultaneously. Their early sequential approach was painfully slow, so they introduced two types of parallelization: the lead agent would spin up 3-5 subagents simultaneously rather than one after another, and the subagents themselves would use multiple tools in parallel.

The results were dramatic: research time dropped by up to 90% for complex queries. This is the AI equivalent of realizing you can have different team members working on different aspects of a project at the same time, rather than waiting for each person to finish before the next one starts.

But here’s where it gets really interesting: Anthropic found that their most effective prompting strategy focused on instilling good heuristics rather than rigid rules. They studied how skilled humans approach research tasks and encoded these strategies directly into their prompts. Decompose difficult questions into smaller tasks. Carefully evaluate source quality. Adjust your approach based on new information. Know when to go deep versus when to go broad.

These aren’t just AI programming principles, they’re fundamental management wisdom that applies whether you’re coordinating artificial agents or human team members.

The system they built now helps users find business opportunities they hadn’t considered, navigate complex healthcare options, and uncover research connections that would have taken days to find manually. But the real insight isn’t about the technology itself, it’s about what effective coordination looks like, regardless of who or what you’re coordinating.

Whether you’re managing a team of people or a system of AI agents, the principles remain remarkably consistent:

• clear objectives
• appropriate scope
• good tool selection
• strategic thinking
• and the wisdom to match your approach to the complexity of the problem at hand

The next time you’re frustrated with delegation not working the way you hoped, remember that even artificial intelligence had to learn these lessons the hard way. Sometimes the best insights about human management come from watching machines try to figure out the same challenges we’ve been wrestling with all along.

If you’re ready to explore how AI can help streamline your own workflows and decision-making processes, let’s connect on LinkedIn and discuss what smart delegation looks like in your world.