Applied AI / Forward Deployed Engineer · Agents & Tool Use · Lesson 3 of 3
Multi-agent, MCP & when NOT to build an agent
Agents are powerful and over-used. Senior judgment is knowing when a simpler design wins.
- Multi-agent systems: specialized agents (planner, researcher, coder, critic) collaborate. More capable for complex work, but more latency, cost, and failure surface. Use only when a single agent genuinely can't cope.
- MCP (Model Context Protocol): an open standard for connecting models to tools/data sources through a uniform interface — think 'USB-C for tools.' Increasingly expected knowledge.
- Workflow vs agent: if the steps are known and fixed, a deterministic workflow (hardcoded chain of LLM calls) is cheaper, faster, and more reliable than an autonomous agent. Use an agent only when the path is dynamic and can't be pre-scripted.
The line that signals maturity"I'd start with the simplest thing that works — often a fixed workflow, not an agent. I add agency only where the task truly needs dynamic decision-making, because every added loop is added latency, cost, and a new way to fail."
Reliability compounds downIf each step is 95% reliable, a 5-step agent is only ~77% reliable end-to-end (0.95^5). More autonomous steps = more compounding error. Constrain scope, add checks, and prefer fewer steps.
◆ Lock it in
- Prefer a deterministic workflow when the path is known; reserve agents for dynamic tasks.
- Multi-agent adds capability but multiplies cost/latency/failure — use sparingly.
- MCP = standard interface for tools/data ('USB-C for tools'). Reliability compounds down across steps.
Feynman drill — say it out loudGive an example task that needs a real agent, and one that should just be a fixed workflow. Justify each.
Step 1 rate your confidence · Step 2 pick your answer
A task has fixed, known steps every time (extract → validate → summarize → email). Best design?
Step 1 — how sure are you?