17. Running a Fault-Injection Exercise for an Agent
A successful happy path proves that the system can run. A controlled failure shows whether it knows where to stop and how to recover.
Latest
18 notesNotes on AI systems
Code, traces, and experiments about agents, language models, and the software systems around them.
16. Building an Agent Regression Suite Before Release
A regression suite is not a rerun of a few demos. It is an executable record of failures the system must not repeat.
09. Agent Permission Is More Than Allow or Deny
Authorization decides whether one principal may perform one action on one resource under specific conditions
07. Are Parallel Tool Calls Actually Faster?
Concurrency shortens only independent waiting time and adds result joining, cancellation, and resource pressure
12. Context, Summaries, and Long-Term Memory Are Different
Longer retention does not create stronger knowledge; unverified memory carries old errors into new tasks
15. Evaluating an Agent Beyond Its Final Answer
A plausible answer does not prove that evidence is real, execution was authorized, or interruption recovery is correct
01. How System Prompts Are Assembled
The model does not receive one prompt. It receives context blocks assembled from different sources for different purposes.
06. How Tool Results Should Return to the Model
The model does not observe local execution; the runtime must carry evidence into the next turn