The short answer
A tool-call contract should define intent, structured inputs, permissions, result states, and recovery behavior.
A tool is not just a function exposed to a model. It is a promise about what will happen when the function is selected.
Names and descriptions should tell the model when a tool is appropriate and when it is not. Ambiguity creates fragile orchestration.
Success, pending, partial completion, and failure should be distinct. Agents can recover when the runtime gives them enough information to choose the next move.
A useful note, once a week
Make room for better signal.
About the author
Rowan studies software tools and the operating habits that make them useful.
More from Rowan Singh →
The winning agent products will not feel like chatbots. They will quietly turn intent into reliable work.
Memory is useful when it helps an agent make fewer repeated decisions without becoming impossible to inspect.
Benchmarks are a starting point. Here is a practical rubric for choosing the right tool stack for a real team.
Keep the good stuff coming
One thoughtful note on AI, agents, and AGI. No noise.
Have a thought? Comments are powered by a lightweight GitHub discussion embed.
