The fastest way to get disappointed with AI is to ask it to improvise inside a system you haven't defined yet.
An agent needs a contract
For an agent to be useful, at least this must exist:
- a clear input,
- a verifiable objective,
- a scope limit,
- a useful output for the next step.
When that's missing, the agent doesn't accelerate anything. It just trades explicit work for diffuse work.
Where it actually pays off
In technical review, artifact generation, context summarization, and repeatable task execution, the return is usually immediate.