AI AGENT / LEARNING & ENGINEERING
Learn to build LLM applications: model interfaces, tools, retrieval, reliable execution, evaluation, and deployment. Progress from concepts to implementation, failure diagnosis, and engineering trade-offs, then test your understanding with questions, tasks, and runnable labs.
RESEARCH NOTES
It focuses on task recovery, execution evaluation and memory management, combined with code and fault experiments.
RUNTIME
Follow checkpoints and leases through human approval, cancellation, and reconciliation of unknown outcomes, with runnable code, crash reproductions, and Java service boundaries.
MEMORY
Inspect schemas, write transactions, retrieval filters, and deletion experiments to see how an old draft can restore information after its source record is deleted.
EVALUATION
Turn success criteria into a test contract using real events and business receipts. Understand the limits of baseline and approval tests, then extend evaluation to semantic quality and cost.
RUN · BREAK · INSPECT
Python 3.10+ · Standard library · Offline by default. Experiments run on your computer or server.
Covers checkpoints, leases, retries, cancellations, idempotent releases, memory invalidations, and manual approvals.
python3 cli.py submit
python3 cli.py run
python3 -m unittest discover -vAfter a crash, where do you resume?Exit the process and verify that submitted steps are not rerun.
After approval, how to proceed?Bind action summaries and versions to prevent expired approvals.
What if the connection fails after dispatch?Verify the actual receipt and do not treat the unknown as unexecuted.
ABOUT THIS NOTEBOOK
This site shares engineering research and experiments written with AI assistance. Explanations of official mechanisms link to primary sources, and examples state their test scope. Content is revised over time; each article shows its update date and published version.