The AI-Agent Stampede

In 2015, "break the monolith" became a mantra in every architecture review. Today the mantra is "go agentic."

The words are different. The stampede is the same.

The 2015 Playbook

Microservices solved a real problem: organisational scale. Independent deploys, team autonomy, engineers not stepping on each other. Netflix's engineering blog became scripture.

Then Docker and Kubernetes made it easy for everyone, and "we're doing microservices" became a hiring signal and a resume line.

Teams with 12 engineers split working monoliths into 45 services.

Many quietly merged them back. The companies that benefited had the organisational maturity to run distributed systems. The rest paid a complexity tax for a problem they never had.

Same Stampede, Different Bill

"Agentic" is now the label everything must carry.

But this time the cost is different. Microservices added deterministic complexity, which you could trace. Agents add non-determinism, token spend and latency on top:

• An orchestration layer to coordinate them
• Retry logic for when one agent's output breaks another's input
• Observability across dozens of live decision points

The complexity moved from the codebase to the runtime, where it is much harder to see.

Three Questions Before You Go Agentic

• Would a simpler automation do the job?
• Is the data these agents will act on clean, accessible and trusted?
• Who owns the outcome when the agent is wrong?

Five agents for every problem multiply cost and erode debuggability and reliability. Not everything needs to go agentic.

The Whiteboard Test

If you can draw the decision tree on a whiteboard before building, it is automation.

If the tree exists only because the model decides it live, case by case, agents earn their place. Think long-running workflows with genuine branching and heavy tool use.

The microservices era taught us that architecture fashion is expensive for those who follow it without the maturity to run it.