MCP Tools Should Separate Facts From Assumptions
Street_Inevitable_77 · reddit · 2026-07-15
In their MCP tool stack, the author divided tools into two categories:
- Fact tools: Only return known, deterministic data from the server, such as current balances or historical records.
- Prediction tools: Only return scenario results based on assumptions, explicitly labeling them as "what-if" scenarios based on growth rates, savings rates, etc.
After implementing this, the model stopped treating "guesses" as certain as facts, naturally using phrases like "under this scenario" instead of jumping to conclusions. The author believes the key is: letting the structure of the return value express confidence, rather than relying solely on tool description text.
They also raised an open question: should a tool that is "90% factual with only one assumed input" be classified on the fact side or the prediction side? The comment section is also debating how to draw the line between deterministic and speculative outputs.
More from coding & agent
- Soft Clamp cuts tool-call overuse in multi-teacher distillation, from 13.7% to 9.0% — antgroup · 2026-07-21
- Agent harness memory loss and compaction are still a major usability problem — adityaag · 2026-07-21
- SpecJudge runs locally on Ollama to pick the right-sized AI model for your project — jokiruiz · 2026-07-21
- A developer maps out six design rules for CLIs that humans and AI agents can both use — yujiezha · 2026-07-21
- A coding-agent skill that forces ADHD-friendly, answer-first output — ayghri · 2026-07-21
- A set of agent skills for CAD, robotics, and hardware design — earthtojake · 2026-07-21