# Agentic Kernel Generation > Source: > Published: 2026-10-19 22:05:17+00:00 # Agentic Kernel Generation - Agentic kernel generation is an automated approach that employs autonomous LLM agents in iterative loops to generate, validate, and optimize computational kernels. - It leverages multi-modal feedback—including static analysis, JIT profiling, and empirical tests—to ensure functional correctness and performance improvement. - These systems support heterogeneous hardware and DSLs, achieving significant speedups and broad operator coverage through adaptive, closed-loop workflows. Agentic kernel generation refers to the automated synthesis, validation, and optimization of computational kernels or system-level primitives using agentic workflows—primarily those driven by LLMs operating in iterative, feedback-driven, and often multi-agent loops. These systems transform kernel enablement from a static human-engineered activity to an adaptive, scalable, and context-sensitive process, targeting diverse hardware and software stacks from AI accelerators to OS subsystems ([Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977), [Wang et al., 31 Jul 2025](https://www.emergentmind.com/papers/2507.23194), [Zhang et al., 23 Oct 2025](https://www.emergentmind.com/papers/2511.01884), [Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236), [Du et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23424), [Dong et al., 19 Oct 2025](https://www.emergentmind.com/papers/2510.16996), [Zhang et al., 19 Nov 2025](https://www.emergentmind.com/papers/2511.15915), [Zheng et al., 1 Sep 2025](https://www.emergentmind.com/papers/2509.01245)). ## 1. Core Concepts and Definitions Agentic kernel generation systems are distinguished by their use of autonomous agents—typically implemented via LLMs—which iteratively generate, assess, and refine kernel implementations. Unlike naive prompt-based or one-shot code generation, these systems use closed feedback loops incorporating both programmatic and empirical checks: - **Agentic loop** : The central workflow, frequently modeled as a[finite-state machine](https://www.emergentmind.com/topics/finite-state-machine-fsm) ([FSM](https://www.emergentmind.com/topics/fast-spatial-memory-fsm) ) or as a search (tree or graph) traversed by cooperative agents. Stages include code generation, static analysis, compilation, hardware execution, and feedback extraction. - **Coverage orientation** : Prioritization of correct functional coverage across large kernel/operator sets, supporting all data types, signature patterns, and argument shapes ([Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977) ). - **Multi-modal feedback** : Integration of static program checks (linting, AST analysis), dynamic runtime profiling (JIT, hardware counters), empirical correctness (test harnesses), and knowledge retrieval from documentation or historical experience ([Zhang et al., 23 Oct 2025](https://www.emergentmind.com/papers/2511.01884) ,[Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236) ). - **Iterative refinement** : Use of[LLM-based agents](https://www.emergentmind.com/topics/llm-based-agents) or subagents specialized for code synthesis, error diagnosis, optimization suggestion, or plan decomposition, operating in feedback loops inspired by human engineering workflows. - **Heterogeneous and cross-platform support** : Compatibility with multiple hardware backends (e.g., NVIDIA, AMD, Meta MTIA, NPUs, CPUs) and diverse kernel[DSLs](https://www.emergentmind.com/topics/domain-specific-languages-dsls) (Triton, CUDA, CuTe, TileLang) ([Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236) ,[Du et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23424) ). ## 2. Architectures and Methodological Taxonomy Several design archetypes for agentic kernel generation have converged in recent work: | System | Architecture | Agents / Submodules | Hardware / DSLs | |---|---|---|---| | TritorX ( [Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977) ) | FSM per operator | LLM generator, Linter, Compiler, Test harness, Log Summarizer | Meta MTIA, Triton | | GEAK ( [Wang et al., 31 Jul 2025](https://www.emergentmind.com/papers/2507.23194) ) | [Multi-agent pipeline](https://www.emergentmind.com/topics/multi-agent-pipeline) | Generator, Evaluator, Reflector, Optimizer | AMD MI300X, Triton | | [CudaForge](https://www.emergentmind.com/topics/cudaforge) ([Zhang et al., 23 Oct 2025](https://www.emergentmind.com/papers/2511.01884) ) | Two-agent (Coder, Judge) | Correction, Optimization (hardware feedback) | CUDA, NVIDIA [GPUs](https://www.emergentmind.com/topics/graphics-processing-units-gpus) | | [KernelEvolve](https://www.emergentmind.com/topics/kernelevolve) ([Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236) ) | Graph search (universal operator) | Node selection, Universal operator, Eval, Retriever | NVIDIA, AMD, MTIA. Triton, CuTe, [MLIR](https://www.emergentmind.com/topics/multilevel-intermediate-representation-mlir) | | AKG ( [Du et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23424) ) | Closed-loop, modular | Designer, Coder, Verifier, [Conductor](https://www.emergentmind.com/topics/conductor) | Triton, CUDA-C, TileLang, CPP | | [STARK](https://www.emergentmind.com/topics/spatial-transcriptomics-via-adaptive-regularization-and-kernels-stark) ([Dong et al., 19 Oct 2025](https://www.emergentmind.com/papers/2510.16996) ) | Tree search, multi-agent | Search controller, Plan agent, Code agent, Debug/Profiler | CUDA | | [AccelOpt](https://www.emergentmind.com/topics/accelopt) ([Zhang et al., 19 Nov 2025](https://www.emergentmind.com/papers/2511.15915) ) | Beam-search loop | Planner, Executor, Summarizer, Memory | AWS Trainium/NKI | | [SchedCP](https://www.emergentmind.com/topics/schedcp) ([Zheng et al., 1 Sep 2025](https://www.emergentmind.com/papers/2509.01245) ) | Multi-agent, decoupled OS | Observation, Planning, Execution, Learning | Linux eBPF, Schedulers | Most implementations structure kernel generation as an iterative process: (1) candidate generation via an LLM (often context-conditioned), (2) static or dynamic formal verification, (3) JIT compilation or hardware execution, and (4) response-driven prompt or memory updates. Architectures range from explicit [FSMs](https://www.emergentmind.com/topics/neural-finite-state-machines-fsms) ([Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977)), beam or tree search ([Zhang et al., 19 Nov 2025](https://www.emergentmind.com/papers/2511.15915), [Dong et al., 19 Oct 2025](https://www.emergentmind.com/papers/2510.16996), [Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236)), to multi-agent modular systems ([Du et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23424), [Wang et al., 31 Jul 2025](https://www.emergentmind.com/papers/2507.23194), [Zhang et al., 23 Oct 2025](https://www.emergentmind.com/papers/2511.01884)). ## 3. Formal Decision Criteria and Optimization Objectives Agentic kernel generation systems formalize correctness and fitness criteria as binary and continuous objectives grounded in hardware-realized execution: - **Lint and static correctness** : Candidate passes if all linter rules yield zero violations: - **Functional correctness** : Operator passes if outputs match a canonical backend within across all relevant test inputs: - **Coverage** : Fraction of operators or benchmarks with complete pass rates; indicates full correctness. - **Performance objectives** : Speedup relative to reference implementation, e.g., TritorX’s fitness: - **Termination** : Completion upon reaching target coverage, improvement stall, or artifact budget exhaustion: Agent selection and expansion often use softmax, -greedy, or [Monte Carlo Tree Search](https://www.emergentmind.com/topics/monte-carlo-tree-search) policies over observed fitness or coverage scores ([Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236), [Dong et al., 19 Oct 2025](https://www.emergentmind.com/papers/2510.16996)). In evaluation, systems report metrics such as median speedup, percent exceeding baseline, pass@K, and per-operator correctness. ## 4. Feedback Mechanisms and Context Management Effective agentic kernel pipelines depend on multi-level and multi-modal feedback, including: - **Static linter/AST analysis** blocking unsafe or “cheating” constructs (e.g., host fallback, recursive ATen calls) ([Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977) ). - **JIT compile/test failures** summarized and filtered for prompt brevity; secondary LLMs often condense error logs ([Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977) ,[Zhang et al., 23 Oct 2025](https://www.emergentmind.com/papers/2511.01884) ). - **Empirical profiling** : Distributed execution on real hardware (FPGA/ASIC/GPU/CPU/NPU), capturing performance counters, occupancy, memory throughput ([Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236) ,[Wang et al., 31 Jul 2025](https://www.emergentmind.com/papers/2507.23194) ,[Zhang et al., 23 Oct 2025](https://www.emergentmind.com/papers/2511.01884) ). - **Retrospective experience/memory** : Explicit archives of slow–fast kernel pairs, with summarizing LLMs to extract transferrable transformations ([Zhang et al., 19 Nov 2025](https://www.emergentmind.com/papers/2511.15915) ). - **Contextual retrieval** :[Retrieval-augmented prompts](https://www.emergentmind.com/topics/retrieval-augmented-prompts) fuse runtime bottlenecks, prior kernel variants, and documentation slices to inform the next iteration ([Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236) ,[Du et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23424) ). [Context management strategies](https://www.emergentmind.com/topics/context-management-strategies) include prompt truncation, focused tokenization (e.g., bottleneck-extracted artifacts only), and dynamic context windows specific to each agent’s role (planning, coding, debugging) ([Dong et al., 19 Oct 2025](https://www.emergentmind.com/papers/2510.16996)). ## 5. Empirical Evaluation and Benchmark Results Scalable agentic kernel systems report the following empirical capabilities: - **Operator and primitive coverage** : TritorX generated correct wrappers for 481/568 PyTorch ATen operators on MTIA (84.7% OpInfo coverage, tests) ([Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977) ). KernelEvolve achieved correctness on 250[KernelBench](https://www.emergentmind.com/topics/kernelbench-cuda) problems and 160 ATen operators across three platforms ([Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236) ). - **Performance** :[Agentic systems](https://www.emergentmind.com/topics/agentic-systems) yield consistent speedups. CudaForge attains a median over PyTorch on diverse GPUs ([Zhang et al., 23 Oct 2025](https://www.emergentmind.com/papers/2511.01884) ). KernelEvolve achieves up to 0 on specific tasks, and AKG reports 1 speedup on Triton-CUDA kernels ([Du et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23424) ,[Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236) ). - **Efficiency and cost** : CudaForge requires only 20.33\sim 26.5$ min wall-clock, substantially below prior agentic baselines ([Zhang et al., 23 Oct 2025](https://www.emergentmind.com/papers/2511.01884) ). - **Robustness & generality** : Complex operator sets, broad datatypes, model-in-the-loop testing (NanoGPT, DLRM, MM1, MM2), and cross-platform adaptability are directly validated ([Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977) ,[Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236) ). - **Ablations** : Removal of critical agents (linter, compilation log summarizer, optimizer) degrades coverage and performance significantly ([Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977) ,[Wang et al., 31 Jul 2025](https://www.emergentmind.com/papers/2507.23194) ). ## 6. Key Design Trade-offs and Future Directions - **Coverage-first vs. performance-first** : Systems like TritorX and KernelEvolve prioritize broad operator support and functional correctness, often deferring fine-grained autotuning. Others (GEAK, AccelOpt) explicitly tune for hardware efficiency post-correctness ([Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977) ,[Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236) ,[Zhang et al., 19 Nov 2025](https://www.emergentmind.com/papers/2511.15915) ). - **FSM vs. fully agentic execution** : FSMs offer stringent control/reproducibility;[agentic architectures](https://www.emergentmind.com/topics/agentic-architectures) with tool-enabling LLMs promise more flexible, adaptive workflows. The field is evolving toward hybrid[agentic orchestration](https://www.emergentmind.com/topics/agentic-orchestration) with tool APIs as first-class interfaces ([Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977) ,[Du et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23424) ). - **Context and memory management** :[Retrieval-augmented prompting](https://www.emergentmind.com/topics/retrieval-augmented-prompting) and lightweight summarization ensure scalability and cost control; explicit[long-term memory](https://www.emergentmind.com/topics/long-term-memory-ltm) or archive-guided planning accelerates convergence and transfers optimization patterns ([Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236) ,[Zhang et al., 19 Nov 2025](https://www.emergentmind.com/papers/2511.15915) ,[Du et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23424) ). - **Extensibility** :[Agentic frameworks](https://www.emergentmind.com/topics/agentic-frameworks) are structured for rapid integration of new hardware backends and DSLs by swapping out DocSpecs, knowledge bases, or hardware constraints in context ([Du et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23424) ,[Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236) ). - **Safety and validation** : Production[agentic deployment](https://www.emergentmind.com/topics/agentic-deployment) mandates strict anti-cheating policies, comprehensive test harnesses, and staged verification to guarantee correctness under all observed usage ([Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977) ,[Zhang et al., 23 Oct 2025](https://www.emergentmind.com/papers/2511.01884) ). Possible directions include reinforcement or [active learning](https://www.emergentmind.com/topics/active-learning-al) for plan selection, [embedding-based retrieval](https://www.emergentmind.com/topics/embedding-based-retrieval-ebr) for optimization memory, multi-platform joint search, and coordinated [agentic optimization](https://www.emergentmind.com/topics/agentic-optimization-aaio) across kernel, OS, and system stack subsystems ([Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236), [Zheng et al., 1 Sep 2025](https://www.emergentmind.com/papers/2509.01245), [Zhang et al., 19 Nov 2025](https://www.emergentmind.com/papers/2511.15915)). ## 7. Broader Context and Philosophical Underpinnings The agentic kernel generation paradigm signifies a transition from monolithic, single-shot code generation toward open-ended, self-adaptive, and robust code synthesis systems, moving beyond traditional AI-hardware co-design cycles ([Hammond et al., 3 Dec 2025](https://www.emergentmind.com/papers/2512.10977), [Liao et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23236), [Du et al., 29 Dec 2025](https://www.emergentmind.com/papers/2512.23424)). The term "agentic kernel" also echoes research in cognitive architectures, where a minimal "functional kernel" enables autonomous emergence of higher-level cognitive functions through reflexive, schema-based self-organization ([Serov, 2022](https://www.emergentmind.com/papers/2207.00822)). This analogy underscores the trajectory of future agentic kernel platforms: to provide the substrate from which both routine and emergent computation can be self-organized and optimized—potentially closing the last-mile gap in hardware–software co-evolution.