Validated tool calls let home-energy agents match 96.7%–98.0% of optimizer savings
MaryamMiradi · x · 2026-07-25
A UK Centre for Computational Science project shows how to build a safe home-energy AI agent around validated tool calls rather than free-form text promises.
Main findings
- The agent translates household requests into hard rules: deadline, duration, power cap, and flexible preferences.
- It pulls live data such as electricity prices, weather, and solar forecasts before choosing a schedule.
- GPT, Gemini, and Claude achieved 100% scheduling success when they used structured tool calls; text-only actions were much less reliable.
- Over a simulated week, the agents captured 96.7%–98.0% of the savings found by a mathematical optimizer, projecting about £1,270/year versus a fixed off-peak timer.
Production lesson
- LLMs should translate intent.
- Tools should compute feasibility.
- A deterministic safety gate should decide whether an action is allowed.
- No validated tool call, no real-world action.
More from coding & agent
- CLI gives models more control over context than badly designed MCP tools — jobergum · 2026-07-25
- A Codex workflow uses a custom “oracle” tool to build ChatGPT Pro access skills — steipete · 2026-07-25
- LLaDA2.2 targets the real bottleneck in multi-turn agents: decode speed — Direct_Band896 · 2026-07-25
- Claude Opus 5 edges out Fable 5 overall, but Fable still leads coding-agent use — Hesamation · 2026-07-25
- Microsoft schedules a free AI Agent Open Hack in Minneapolis on July 29 — WirelessLife · 2026-07-25
- Opus 5 guide says to drop explicit verification prompts for agentic coding — cedric_chee · 2026-07-25