Show & Tell: What did the agent understand?
Show the agent how you do the work, explain the decisions, and check what it understood.
When someone new joins, we often ask a colleague to show them the work. We go through an example, answer questions, and see what they understood before leaving them to handle the next one.
This is how we want people to teach an agent, too. In Brackett, we call it Show & Tell.
Teach it while doing the work
You demonstrate the process in your browser and explain the decisions out loud. Show & Tell captures the actions, what was on screen, and your narration together. The agent can ask about a gap while you are still there to explain it.
If you say, “I do this for each row,” the useful question may be, “Just today’s rows, or every row in the sheet?” If you usually message the team afterward, it needs to know which team and where. These are small questions, but the answers change the work.
The person teaching knows where the usual rule stops applying and when to ask someone else. We want to keep that knowledge with the demonstration, so the next agent does not have to guess why a step happened.
Show how the records connect
A supplier’s part-change notice may use a different part number from the one in your own system. You show the agent the notice, look up the supplier’s number in a mapping sheet, and then check the matching material in SAP.
The lesson includes the relationship between those records and the reason for checking them. Finding a number in the notice is only the beginning; we need to know which material in our business it refers to.
After the demonstration, the agent lays out its understanding of the part-change process for review.
Show & Tell is part of how we build an automation in Brackett. The person who knows the process can review the captured lesson and correct what the agent misunderstood before building on it.
Checking what it understood
We built ShowTellArena to examine what products understand after a demonstration. Brackett, Claude, and Codex learn the same scenario through their own teaching interfaces, then answer the same questions. Our early pilot showed differences in both comprehension scores and whether attempts reached a scored answer. The results, examples, and evaluation details are on the research site.
This connects to learning in the Agent Effectiveness Index. Show & Tell provides a way to teach; AEI goes further to review performance on new work, retention, and changed rules. The arena’s immediate comprehension quiz is reported separately from AEI’s Learning Persistence score.
For other agent builders, too
Show & Tell also has an MCP connector. MCP, the Model Context Protocol, lets an agent request a demonstration and read the resulting actions, narration, and screenshots.
We plan to open this to all agent builders later, so their agents can ask a person to show them the work, too. The person already knows how to do it. We want them to be able to teach it in the course of doing their job.