Back to AI Research

AI Research

AtumAI: A Principled Framework for Agentic Generati... | AI Research

Key Takeaways

  • AtumAI is a framework designed to automate the creation of datacenter control-plane policies—the software rules that manage hardware resources like CPU, memo...
  • The efficiency of a datacenter rests on its control plane policies.
  • Designing these policies is increasingly hard: the hardware-software stack grows fast, the design space is vast and interdependent, and prototyping a single policy takes months.
  • Agentic AI promises to automate this search.
  • Off the shelf, however, it falls short on three fronts.
Paper AbstractExpand

The efficiency of a datacenter rests on its control plane policies. Designing these policies is increasingly hard: the hardware-software stack grows fast, the design space is vast and interdependent, and prototyping a single policy takes months. Agentic AI promises to automate this search. Off the shelf, however, it falls short on three fronts. It is not formal: with no structured, searchable statement of the problem, the search has little structure to exploit and hard constraints are not guaranteed. It is not transferable: each task is solved from scratch, so nothing learned on one task carries to the next. Finally, it is not systematic: relying on the LLM as the sole source of candidates, it explores a narrow slice of the design space and settles into local optima. We introduce AtumAI, a framework that generates datacenter control-plane policies with agentic AI, making the process formal, transferable, and systematic. From a goal stated in plain language, AtumAI autonomously proposes, tests, and refines candidate policies until one satisfies the request. It does so through two components. The Datacenter Task Compiler automates problem formulation: it compiles the request into a formal, machine-checkable, and searchable specification of the task's objectives, constraints, decision variables, and evaluation methodology. The Evolutionary Design Discovery Loop then searches this specification, expanding the search beyond the LLM itself via a diffusion model, an evolutionary algorithm, and a surrogate model. Together, they reduce onboarding a new task from months of engineering to writing its description. We evaluate AtumAI on three control-plane tasks with distinct problem scopes, design spaces, and trade-offs: workload placement, resource scaling, and power management. Across all tasks, the policies generated by AtumAI consistently outperform expert-engineered baselines.

AtumAI is a framework designed to automate the creation of datacenter control-plane policies—the software rules that manage hardware resources like CPU, memory, and power. While human experts currently design these policies, the increasing complexity of modern hardware and software makes manual development slow and prone to missed efficiency opportunities. AtumAI uses agentic AI to autonomously propose, test, and refine these policies based on plain-language goals, reducing the development timeline from months to hours.

Addressing Limitations in Agentic Design

The authors, Qiushi Lin et al., argue that off-the-shelf agentic systems are insufficient for datacenter control-plane tasks because they lack three critical properties:

  • Formality: Existing systems often lack a structured, machine-checkable way to define objectives and constraints, which can lead to policies that fail to meet hard requirements.

  • Transferability: Most AI-driven solutions are built for a single, bespoke task, meaning knowledge gained from one project cannot be reused for another.

  • Systematic Search: Relying solely on a Large Language Model (LLM) for candidate generation often results in biased, narrow searches that get trapped in local optima.

How AtumAI Works

AtumAI addresses these gaps through two primary components:

  1. The Datacenter Task Compiler: This front-end component translates a natural-language request into a formal, machine-checkable intermediate representation (IR). This IR explicitly defines decision variables, objectives, constraints, and evaluation methodologies. By grounding these specifications in real workload and platform data, the compiler ensures the resulting policy is tailored to the specific environment. 2. The Evolutionary Design Discovery Loop: This back-end component performs the search. To move beyond the limitations of LLM-only generation, it employs a diffusion model to explore structurally distinct designs, an evolutionary algorithm to tune parameters, and a surrogate model to filter out poor candidates before they undergo costly evaluation.

Performance and Results

The researchers evaluated AtumAI across three distinct control-plane tasks:

  • Workload Placement: AtumAI improved placement success rates by 17% and scheduler throughput by 8% compared to expert-engineered baselines.

  • Resource Scaling: The framework achieved a 24% improvement in cost efficiency while maintaining service-level objective (SLO) violations at 1.3%.

  • Power Management: AtumAI reduced power consumption by 21% while simultaneously increasing throughput by 17%.

Franklin Analysis

The evidence suggests that AtumAI’s primary innovation is the introduction of a formal, reusable intermediate representation (IR) that decouples the policy-generation process from specific domain tasks. By moving from a "bespoke" model to a shared library of control knowledge, the framework allows for the accumulation of expertise across different datacenter problems. The use of a multi-modal search strategy—combining LLMs, diffusion models, and evolutionary algorithms—appears to be the key factor in overcoming the narrow search biases typically associated with pure LLM-based agentic systems.

Comments (0)

No comments yet

Be the first to share your thoughts!