Andrii Bidochko

• Updated: July 11, 2026
• 7 min read

Agentic Neural Architecture Search

Direct Answer

Agentic Neural Architecture Search (AgentNAS) introduces a hybrid pipeline that lets a large language model (LLM) draft a high‑quality seed neural network, then automatically converts that seed into a “slotted” architecture—a scaffold with interchangeable modules that defines a bounded, task‑specific search space for conventional NAS to explore. This division of labor dramatically reduces the need for hand‑crafted search spaces while still delivering state‑of‑the‑art performance across a wide variety of vision and regression tasks.

Background: Why This Problem Is Hard

Neural Architecture Search has become a cornerstone of modern AI development because it can uncover model topologies that outperform human‑designed baselines. However, the process still hinges on two fragile pillars:

• Manual search‑space engineering: Researchers must encode domain knowledge into a fixed set of allowable operations, layer types, and connectivity patterns. This step is time‑consuming, error‑prone, and must be redone for every new dataset or modality.
• Computational expense: Even with recent efficiency tricks (weight sharing, one‑shot methods), exploring a large combinatorial space remains costly, especially when the space is poorly aligned with the target task.

At the same time, LLMs have demonstrated an uncanny ability to generate plausible code snippets, model definitions, and even entire research proposals. Yet, their outputs are stochastic and lack systematic optimization; a single LLM sample rarely matches the performance of a dedicated NAS run. The open question is how to combine the creative breadth of LLMs with the rigorous search capabilities of NAS without re‑introducing manual bottlenecks.

What the Researchers Propose

The authors of the Agentic Neural Architecture Search paper propose a three‑phase framework called AgentNAS. The core ideas are:

1. LLM‑driven seed generation: An LLM is prompted to produce a complete neural network description for the target task. The seed is already competitive with many published baselines.
2. Slotted architecture conversion: The seed is parsed into a scaffold where each logical component (e.g., feature extractor, classifier head, up‑sampling block) becomes a named “slot.” Each slot can host a set of interchangeable modules drawn from a curated library.
3. NAS‑driven slot recombination: A conventional NAS algorithm searches over the combinatorial space defined by the slots, selecting the best module for each slot and optionally tuning hyper‑parameters. Because the space is now task‑specific and bounded, the search is both efficient and expressive.

In essence, the LLM supplies creativity, while the NAS engine supplies rigor. The two agents operate sequentially but are evaluated independently, allowing the authors to quantify each contribution.

How It Works in Practice

The AgentNAS pipeline can be visualized as a linear workflow with clear hand‑offs:

1. Prompt design: Engineers craft a concise natural‑language prompt that describes the dataset, performance metric, and any resource constraints (e.g., FLOPs budget).
2. LLM inference: The chosen LLM (e.g., GPT‑4, Claude, or a domain‑specific model) returns a full model definition in a standard framework (PyTorch, TensorFlow). This definition includes layer types, connections, and optional hyper‑parameters.
3. Slot extraction: A parser scans the LLM output, identifies logical groups (convolutional block, attention module, etc.), and replaces each group with a placeholder slot name (e.g., FeatureSlot1, HeadSlot).
4. Module library construction: For each slot, a curated set of candidate modules is assembled. The library can be drawn from existing NAS benchmarks (NAS‑Bench‑360) or from publicly available model zoo components.
5. NAS search: A search algorithm (e.g., evolutionary search, Bayesian optimization) explores the Cartesian product of slot‑module choices, evaluating each candidate on a validation split. Because the search space is limited to the slots, the algorithm converges quickly.
6. Final model selection & training: The best configuration is instantiated, fully trained on the training set, and evaluated on the test set.

What distinguishes this approach from prior work is the automatic generation of a bounded search space directly from an LLM‑produced architecture. No human needs to hand‑craft the space, and the NAS engine never has to wander through irrelevant regions of the design universe.

Evaluation & Results

The authors benchmarked AgentNAS on 17 diverse tasks spanning image classification, dense regression, semantic segmentation, and multi‑label tagging. The datasets were drawn from two sources:

• NAS‑Bench‑360: A curated suite of vision and tabular tasks with publicly available baselines.
• Unseen NAS: Newly introduced tasks that have not been used in prior NAS literature, ensuring a fair test of generalization.

Key findings include:

• On 11 out of 17 tasks, AgentNAS set a new state‑of‑the‑art record, surpassing both handcrafted expert designs and prior NAS methods.
• The LLM‑generated seed alone already outperformed published baselines on the majority of tasks, confirming that LLMs can produce strong initial designs without any search.
• NAS refinement added an average boost of 2–4% absolute accuracy (or equivalent metric improvement) on top of the seed, demonstrating that combinatorial recombination across slots yields gains that single LLM samples cannot replicate.
• Ablation studies across three LLMs of varying capability (small, medium, large) showed consistent patterns: stronger LLMs produce better seeds, but even modest LLMs benefit substantially from the subsequent slot‑search phase.

These results validate the hypothesis that a well‑structured division of labor can capture the best of both worlds: creative, open‑ended generation from language models and systematic, performance‑driven optimization from NAS.

Why This Matters for AI Systems and Agents

For practitioners building AI agents, the AgentNAS paradigm offers several practical advantages:

• Rapid prototyping: An engineer can obtain a competitive model in minutes by prompting an LLM, then let the NAS component polish it automatically. This shortens the iteration loop dramatically.
• Reduced expertise barrier: Teams without deep NAS experience no longer need to design search spaces manually. The “slot” abstraction abstracts away low‑level architectural decisions.
• Scalable orchestration: Because the pipeline is modular, it can be embedded into existing agent orchestration frameworks. For example, the Workflow automation studio can trigger the LLM step, monitor the slot extraction, and dispatch the NAS search as a background job.
• Resource‑aware design: By constraining the module library per slot, organizations can enforce hardware budgets (e.g., mobile‑friendly convolutions) while still exploring a rich design space.
• Continuous improvement: As newer LLMs become available, the seed generation step can be upgraded without re‑engineering the NAS component, ensuring a future‑proof pipeline.

In short, AgentNAS turns architecture design into a serviceable component of an AI agent’s toolbox, aligning with the broader trend of “agentic AI” where autonomous modules collaborate to solve complex problems.

What Comes Next

While AgentNAS marks a significant step forward, several open challenges remain:

• Dynamic slot libraries: Current implementations rely on a static set of candidate modules. Future work could let the NAS engine synthesize new modules on the fly, further expanding the search horizon.
• Cross‑modal transfer: Extending the slotted architecture concept to multimodal tasks (e.g., vision‑language models) will require slot definitions that span heterogeneous data streams.
• Feedback loops: Integrating performance feedback into the LLM prompting stage could create a true closed‑loop system where the LLM iteratively refines its seed based on NAS outcomes.
• Robustness and safety: Automatically generated architectures must be vetted for adversarial robustness, fairness, and compliance—areas that are still largely manual.

Addressing these gaps will likely involve tighter coupling between language models and differentiable architecture generators, as well as richer evaluation suites. Organizations interested in experimenting with agentic model design can start by exploring the UBOS platform overview, which provides a sandbox for integrating LLMs, custom module libraries, and automated search pipelines.

Moreover, the same slotted‑architecture idea can be repurposed for other AI‑centric workflows, such as automated prompt engineering or tool‑selection for autonomous agents. By treating any composable system as a set of interchangeable slots, developers can leverage the same “design‑then‑search” pattern across the entire AI stack.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.