A model can pick the wrong tool, attach a real quote to a false claim, or overwrite the chart a user just corrected. These notes are about those specific problems. Each stands on its own; there is no required reading order.
The examples drawn from professional work are generalized architecture notes, with illustrative tool names and payloads. The MCP article comes from my own arcade research and includes a dated record of what was tested.
Pick the problem you have
| Problem | Start here | The decision |
|---|---|---|
| Too many similar tools | Your Agent Has Every Tool. It Still Picks the Wrong One. | Retrieve a short list, then fetch one schema; measure the retrieval failures this introduces. |
| Plausible citations that do not support the answer | The Citation Is Real. The Answer Is Wrong. | Check source location, structure, and factual support separately. |
| A new model looks better in a demo | Your New AI Model Looks Smarter. It Might Be a Downgrade. | Compare against the current system on the same cases, with release criteria written first. |
| An agent keeps undoing human edits | Your AI Changed the Colors. Your Edits Disappeared. | Give artifacts an identity, track revisions, and make approval apply to the revision reviewed. |
A working example from the lab
Your AI Read the Repo. It Still Missed the Blocker. connects these concerns to real repositories: bounded retrieval, pinned evidence, explicit blockers, and tests that reject unsafe or misleading requests.
The September 2 record covers 21 read-only tools and 144/144 harness tests. An earlier twelve-tool trial returned 8,244 characters from a 43,547-character source set. That comparison describes one retrieval task; it does not establish token savings, better port quality, or current support across every agent client. The article keeps those open questions alongside the results.
For the human reason behind some of this tooling, read Alan Is Gone. I'm Trying to Keep His Fighting Style Alive.. For the broader portfolio, see Selected Work: The Games I Couldn't Leave Alone.