Back to the journal

Better AI agents show their work

Why editable intermediate work can make an AI assistant easier to guide, verify, and improve.

AI agents / Product engineering

An assistant becomes easier to collaborate with when it exposes the work between a request and its answer. A query, a small script, or a visible set of source records gives the user something concrete to inspect and correct.

The original essay illustrated this with a project-status report assembled from several workplace tools. If the assistant selects the wrong channel or date range, a visible intermediate step lets the user identify the mistake before accepting the summary.

Make correction part of the interface

The design argument is straightforward: show which information was used and how it was transformed. Let the user revise an assumption and continue from that point. Treat example scenarios as illustrations, not measured productivity results.

Keep capability and control together

MCP means Model Context Protocol. It does not require a fixed linear workflow, and code execution can complement exposed tools. Neither approach removes the need for scoped permissions and review of consequential operations.

An isolated execution environment can reduce risk, but it cannot guarantee harmless behaviour. The useful goal is an assistant whose actions and intermediate results remain understandable enough for a person to guide and verify.

About the author

Timur Isachenko is an AI CTO and Principal Solution Architect working on AI-native products and distributed systems. Explore his engineering projects and technical books, or read more about his experience.

Keep following your curiosity.

Back to the journal.
Better AI agents show their work | Timur Isachenko