Before it
An agent that failed a task simply failed and started over.
The problem
Improving from failure normally means gradient updates, which is impossible at inference time.
What it introduced
Have the agent write a verbal post-mortem of its failure and keep it in memory for the next attempt.
Why it matters
Learning from experience without touching the weights, which became a standard pattern in agent design.
What modern models inherit
Episodic self-reflection as an agent's cheap substitute for training.
What came next
Memory and self-critique layers in agent frameworks and coding agents.