Posts newest first

1. 09.20
   The record your AI search missed: start with the indexed text

   The cause memo a staff member wrote was not in the search text. Even after merging the text, sending only the first 200 characters can drop that sentence. Follow one record’s evidence to find the step to fix, and check the difference in a small lab.

   Building agents11 min read

   Text first
2. 09.10
   Merging changes that coding agents built separately

   When the approval API and the screen expect different error responses, each side’s tests pass and the connected feature still breaks. A worked example of splitting the work, agreeing on the response, and ordering review and integration checks.

   Coding agents12 min read

   Answer key
3. 09.08
   From a chat message to an action on the actual work screen

   One line, "open the work request list", travels through the model’s tool choice, the server’s receipt acknowledgement, AG-UI events and React’s navigation function. I followed a single screen identifier along that path. The model picks the screen and the browser opens it, and the bot’s "I have moved you there" rests on nothing more than the server’s acknowledgement.

   Building agentsADK Java series Part 316 min read

   Received only
4. 09.03
   Gemini function calling in Java: how much you end up doing yourself

   Gemini selects a function and its arguments; your Java application owns the rest. Here is the manual loop, its silent cross-layer failures, and why I chose Google ADK Java with AG-UI.

   Building agentsADK Java series Part 214 min read

   Hand-off
5. 08.17
   How one translation request became an internal LLM platform

   A translation button, asked for in passing. One genai call would have finished it in a day. Instead I built the layer that keeps prompts, instructions, and tools out of the code and on a screen.

   Building agentsADK Java series Part 18 min read

   Groundwork