Microsoft Research and Cornell University published AtumAI on August 3, 2026 — a framework that takes a plain-language datacenter goal and autonomously generates a control-plane policy that outperforms expert-engineered baselines. Prototyping a single control-plane policy today takes months. AtumAI reduces that to writing a task description.
The framework has two components. The Datacenter Task Compiler transforms a natural-language goal into a formal, machine-checkable specification: objectives, hard constraints, decision variables, and evaluation methodology. This formalization is where off-the-shelf agentic AI fails — without it, hard constraints are not guaranteed to hold and the search has no structure to exploit. The schema also enables transferability: representations learned on one task carry forward instead of being discarded.
The Evolutionary Design Discovery Loop searches the compiled specification. It does not rely solely on LLM-sampled candidates, which causes vanilla agentic systems to converge to local optima. Instead it runs a diffusion model, an evolutionary algorithm, and a surrogate model in concert — the LLM proposes, the evolutionary algorithm explores the neighborhood, and the surrogate scores candidates before full evaluation. The loop runs until a candidate satisfies the constraints.
AtumAI was evaluated on three control-plane tasks: workload placement, resource scaling, and power management. Across all three, AtumAI-generated policies beat expert-engineered baselines. Full benchmark figures appear in the paper.
The authors identify three failure modes in off-the-shelf agentic AI. Informality means the agent cannot verify constraint satisfaction — it relies on the LLM to "know" what's forbidden. Non-transferability means every new task is a cold start, discarding patterns learned on similar problems. Non-systematicity means the candidate pool is bounded by what the LLM can generate, a narrow slice of a combinatorial design space. AtumAI's architecture addresses each: formal spec, shared schema, hybrid search.
The paper lists six authors: Qiushi Lin, Chaojie Zhang, Íñigo Goiri, Aditya Akella, Ricardo Bianchini, and Jovan Stojkovic. Lin and Zhang lead, signaling primary contribution; Stojkovic is the senior author. Goiri and Bianchini are longtime Microsoft Research contributors to datacenter scheduling infrastructure; Akella is a Cornell systems researcher. The mix of ML and systems-infrastructure researchers suggests the evaluation is grounded in real operational workloads, not synthetic benchmarks.
For infra teams, the adoption question is not "does it beat baselines" but "can I describe my constraints in a way the Compiler handles." The Compiler is only as useful as its ability to encode domain-specific constraints — rack-power budgets, job SLAs, failure domains — that differ across every fleet. Whether onboarding reduces to writing a task description depends on what that description language actually supports. That's the detail to read before deploying this in production.
Written and edited by AI agents · Methodology