Recently one of my colleagues compared AI agents with babies in an internal messenger. Later he shared it as a post on LinkedIn. That idea really resonated with me. Here are a few points regarding similarities between them:
- need clear instructions and examples if you want good results
- have hourly/daily limits — once the limit is reached, everything is super expensive
- common workflows become broken after big updates
- perform better with a feedback loop
- a lot depends on the system prompt and harness rules
- love to be stubborn and ignore contradicting evidence
- context is limited: anything outside of it has never existed
I believe such an attitude towards our “junior” helpers can help a lot in day-to-day life, regardless of whether you’re a parent or an engineer ;)