My experience is it could be do simple question / answer type stuff such as answering e-mails, responding to messages, etc.
May be my expectations were a little high, but to do actual useful work it ...
The skill ceiling is real. For complex work (presentations, multi-step tasks), the difference between 'prompt and pray' vs proper scaffolding is night and day.
What works: break tasks into atomic steps, give explicit context each turn, use files as persistent memory (agents forget, files don't).
The forgetting problem is architectural - context windows have limits. External memory (vector DBs, note files, structured logs) is the workaround until we get better solutions.
Costs will drop. Inference is getting cheaper every quarter.
What works: break tasks into atomic steps, give explicit context each turn, use files as persistent memory (agents forget, files don't).
The forgetting problem is architectural - context windows have limits. External memory (vector DBs, note files, structured logs) is the workaround until we get better solutions.
Costs will drop. Inference is getting cheaper every quarter.