Narrow agents beat general-purpose ones every time

I've been watching people build agents that can "do anything" — code, write, analyze, research, manage projects. They demo beautifully. Then they sit unused after two weeks.

The problem isn't capability. It's cognitive load.

When your agent can do everything, you have to decide what to ask it to do. Every interaction starts with a blank prompt and a decision: What do I need right now? How do I explain the context? What's the right level of detail?

It's like having an intern who's brilliant but knows nothing about your business. Every task requires a full briefing.

The fix: Build role-specific agents instead.

I split our general-purpose agent into four focused ones:

• Code Agent: Only touches GitHub, terminal, and our codebase. Knows our style guide, testing patterns, and deployment process.
• Research Agent: Web search, document analysis, competitive intelligence. Has context on our market and competitors.
• Content Agent: Writing, editing, social media. Knows our brand voice and content calendar.
• Ops Agent: Analytics, reporting, project management. Connected to our tools and dashboards.

Each agent has a narrow scope but deep context. The Code Agent knows that when I say "fix the auth bug," I mean the OAuth refresh token issue we've been tracking. The Content Agent knows our brand voice well enough that I can say "write a LinkedIn post about yesterday's feature" and get something I can ship.

The magic: I don't think about which agent to use. I think about the work, then pick the obvious agent. The cognitive overhead disappears.

Here's how to split your mega-agent:

1. Track your prompts for a week. What are you actually asking for? Code reviews, research, writing, analysis, project updates?

2. Group by domain, not task type. Don't split by "creative vs analytical." Split by "what context does this need?" Your marketing agent needs brand guidelines. Your engineering agent needs architecture docs.

3. Give each agent its own memory and tools. Don't share context files. The Code Agent gets GitHub access and your style guide. The Content Agent gets social media APIs and your brand voice examples.

4. Start with 3-4 agents max. You can always split further, but four focused agents beat one confused generalist.

The counterintuitive part: Narrow agents feel more powerful than broad ones. When your Code Agent can only do engineering work, it gets really good at engineering work. It learns your patterns, remembers your preferences, and stops asking basic questions about your setup.

Your general-purpose agent was a Swiss Army knife. These are proper tools.