Wink Pings

Reviewing Agent Sessions Every Friday: The Places Where Humans Have to Step In Are Assets

A habit for improving agent performance: every Friday, read through all agent sessions on your machine, find the places where humans had to step in, and turn them into long-lasting environmental assets. Next time a similar task comes up, the agent can do it right on its own.

Ricko Yu shared a technique for improving agent performance:

Every Friday, read through all agent sessions on your machine, find the places where humans had to step in, and turn those places into long-lasting environmental assets—hooks, scripts, skills, conventions, memory.

Specifically: open the session records and focus on where you manually changed the output, added parameters, or corrected the direction. These places are the agent's knowledge gaps. Turn them into automated assets.

The patterns of agent errors are often very stable. The same pitfall it falls into this time, it will fall into again next week. The human remembers, but the agent doesn't. Reviewing sessions is about transcribing human memory into the agent's long-term memory.

And this habit doesn't require any changes to the model or retraining—it relies only on the accumulation of environmental assets.

Next time a similar task comes up, the agent can do it right on its own.

Original post image:

image

The value of this habit lies in treating every human intervention as a requirements analysis: wherever a fallback, correction, or completion is needed, transform that place into an environmental asset, clearing obstacles in similar tasks for the agent in advance.

In the replies, Ricko Yu added that most of the observation metrics in his scenario improved.

image

If you don't want to keep cleaning up after the agent, first let it see how you clean up every time.

发布时间: 2026-11-20 19:31