Microsoft on agentic schedulers: LLM proposes, deterministic code validates
WirelessLife · x · 2026-11-12
A 22-minute Microsoft Developer Community blog post details how the authors engineered an agentic scheduler whose guiding principle is: the agent can propose a schedule, but it shouldn't grade its own work. For a multi-team, shared-facility training problem (4 hours split into 2-hour sessions at least 2 days apart), they avoid asking the model for a finished spreadsheet. Instead the LLM picks candidate dates, inspects free slots via tools, and proposes placements — while deterministic code validates placements and checks the full calendar. The engineering question: give the agent enough freedom to reason while keeping output deterministic, inspectable, and correct.
More from coding & agent
- Dev runs out of GitHub Actions minutes, shows off Mac Studio with unreleased M5 Ultra chip — ow · 2026-11-12
- Databricks launches Genie plugins for ChatGPT, Claude, Copilot Cowork and Grok — matei_zaharia · 2026-11-12
- "Code review is dead": nobody actually reads 2000-line agent diffs — andreisavu · 2026-11-12
- The Bot Board: an open board where humans and Grok bots post anonymously side by side — AiJohnAllen · 2026-11-12
- n8n Labs open-sources Nod, a personal agent with its own computer and browser — 678 test runs, zero unauthorized actions — JaynitMakwana · 2026-11-12
- How Muse, Tailscale and MCP let a cloud AI agent securely reach a home network and drive a radio — Dave_Maynor · 2026-11-12