Focus Compounds - Why Specialized AI Agents Outperform Generalists
Focus Compounds in AI Agent Architecture
There is a temptation to build one agent that handles everything - email, calendar, CRM, code, browser, files. The logic seems sound: one agent with full context should make better decisions than five agents with partial context.
In practice, the opposite happens. A focused agent that handles only CRM updates becomes excellent at CRM updates. A general agent that handles CRM plus nine other things becomes mediocre at all of them.
Why Generalist Agents Underperform
The problem is not capability - it is context allocation. A generalist agent spends significant context window space on:
- System prompts for ten different domains
- Tool definitions it will not use in the current task
- Memory entries from unrelated workflows
- Error handling for scenarios outside the current scope
By the time the agent gets to the actual task, a meaningful chunk of its context budget is already consumed by irrelevant information.
The Compounding Effect of Focus
A specialized agent benefits from compounding in three ways:
- Better system prompts - all instructions target one domain, making them more precise
- Cleaner memory - every memory entry is relevant to the task at hand
- Refined tool usage - the agent gets better at using a small set of tools expertly
Over time, a focused CRM agent learns the specific quirks of your CRM, the patterns in your data entry, and the edge cases that cause errors. A generalist agent never develops this depth.
The Right Architecture
Build small, focused agents and orchestrate them. An orchestrator routes tasks to the right specialist. Each specialist maintains its own memory, tools, and system prompt.
This mirrors how effective teams work. You do not hire one person to do sales, engineering, design, and accounting. You hire specialists who coordinate.
When to Generalize
The exception is personal assistants where cross-domain context genuinely matters - "schedule a meeting with the person I just emailed about the contract we reviewed yesterday." Those workflows benefit from unified context. Everything else benefits from focus.
Fazm is an open source macOS AI agent. Open source on GitHub.