Topic
Design practice for shaping the functions exposed to AI agents: how they're named, scoped, and composed, with clear, unambiguous tool outputs.
An AWS waste scanner now exposes five read-only MCP tools that hand back totals, counts and breakdowns its engine computed. The findings themselves stay in a report file on disk, so the model has nothing to sum.
Reality
A dev.to comparison scores MCP tool design at 13 website builders and content systems against eight criteria. The two designs it walks through in full both take deletion out of the flat tool list and put it behind its own gate.
Reality
A Thai-language walkthrough estimates 15 to 20 MCP calls to build a ceramics studio site on WordPress.com. Counting each documented operation, including one status update per item to publish, gets to 22.
Reality
The agent never reached for the optional lsp tool that carried diagnostics, so SolonCode moved error feedback onto write and edit, with a document version check so an old diagnosis cannot be read as a verdict on the new edit.
Reality
The fixture fallback in a five-agent hackathon build was honest and clearly labelled, and it carried every screen of a deployment whose browser requests never once got through. Good degradation buys long silences.
Reality
One team's account of wiring government open data into agent-callable evidence: the failure mode is a valid response that is only page one, and an agent that reads a cap as "no records".
Reality
JetBrains traced a model through fifteen C# refactoring tasks and found it simulating structure with sed, git and the compiler. Wiring in Rider's real engine cut median time by 83%.
Reality
A dev.to writeup argues protocol conformance tells you nothing about whether a model can finish work with your tools. The tell is a wrong answer built from calls that all succeeded.
Reality
A dev.to series entry on the harness tool layer argues fifty bespoke tools are worse than read_file and bash. The usable test is whether one tool already expresses another.
Reality