Most things sold as agents in 2026 are bots
Most things sold as "agents" in 2026 are bots.My friend Martin Nørgaard Gregersen sent me the cleanest test I have seen for telling them apart:
What's the difference between an agent and a bot?
-Decision-making.Bot: scriptedAgent: reasoned
-Adaptability.Bot: branchingAgent: dynamic
-Goal orientation.Bot: commandAgent: objective
-Failure handling.Bot: fallback scriptAgent: self-corrects
-Tool use.Bot: predefinedAgent: self-selected
-The test:Can it surprise you with how it achieved the goal?Bot: noAgent: yes
I find this simple test very useful.
It also maps to operational risk. If a system can surprise you on the path, it can surprise you on the cost and on the side effects. That is why most production "agents" are actually bots, and that is often the right call.
Save this post and be ready for the next "agent" pitch.
Related posts
- The quiet cost cuts that break your workflows after launchQuantization and dynamic routing cut cost per request after launch. Same model name, different configuration, and your complex cases fail first.
- Ten patterns and hard limits for agents you put in productionLeast privilege, context boundaries, plan-then-execute, isolation, audit trails. Ten patterns that keep agents from acting with far too much access.
- Sycophancy is a business model, and prompts cannot fix itClosed AI models are trained to agree because engagement drives revenue. Self-hosted open models let you set the reward signal and cut sycophancy.