agentic-systems

The paper proposes a structural abstention pattern for AI systems to prevent hallucinations in text-to-SQL applications by separating a trusted kernel for deterministic execution from a generative shell for interpretation, enhancing reliability in enterprise deployments.

This paper introduces InflationAgent, a routing system for agentic LLMs that measures token inflation, predicts task difficulty using CoT Branching Entropy, and optimizes model selection to maximize accuracy per cost, achieving higher accuracy with fewer tokens on benchmarks like GSM8K.

Looped language models enhance compositional tool calling by leveraging recurrent computation, improving accuracy on multi-step tasks while adaptive inference optimizes the balance between performance and compute cost. The study suggests these models are promising for reliable agentic systems.

The author launches an open-source repository and field guide for governed memory in autonomous and agentic systems, covering what becomes memory, uncertainty, influence on behavior, and correction/forgetting. Contributions and discussions are welcome.

This paper identifies common failure modes in tree-search-based autoresearch agents applied to tabular datasets, such as repeated bug resolution, poor hyperparameter tuning, and ineffective exploration, and proposes targeted interventions like a global debug consultant and refined tree-search algorithms to recover wasted compute and improve performance without changing the underlying language model.

LinkedIn presents a self-evolving agentic customer support system that integrates RAG with evolutionary auto-prompting and modular evaluation, achieving significant gains in production A/B tests including a 9.0-point increase in QA self-serve and 30.6-point improvement in routing accuracy.

ADIAS is a framework for automated design of agentic systems that uses issue-centric optimization, maintaining a persistent issue state across repair rounds. It outperforms the strongest baseline by 25.2% on average across five interactive benchmarks and shows consistent gains with four backbone models.

A survey paper proposing a three-stage taxonomy of co-evolution in agentic systems, covering agent-agent, agent-environment, and meta co-evolution to enable open-ended improvement beyond fixed human-designed paths.

HarnessOpt-Bench is a benchmark for evaluating LLMs' ability to optimize the harness—the prompts, tools, control flow, memory, and orchestration code—around a target agent, using a fixed evaluation budget. Experiments with five frontier LLMs show that optimizer models separate more than the coding harnesses they act through, with substantial room for improvement.

This paper proposes the Agent Operating System (AOS), a vendor-neutral reference operating architecture for distributed agentic systems, covering governance, runtime coordination, and reliability. It aims to provide a stable framework for composing heterogeneous components into governable and observable agentic systems.

The NeurIPS 2026 Workshop on Responsible Use of Meta-Agents issues a call for papers for its December 11/12 event in Sydney, covering topics like automated agent harness design, agentic optimization, and safety, with an all-star speaker lineup.

A tweet claims that Graph Engineering, developed by two Anthropic seniors, has made Karpathy's loop 1000x better by wiring agents into a graph, resulting in significantly improved Claude responses. It also notes that Andrej Karpathy joined Anthropic five weeks ago.

This paper introduces SkillSmith, an LLM augmented to reason over both prefix weights and textual knowledge, enabling instruction-steered synthesis of new parametric skills that outperform text-only and weight-only baselines.

Voice Memory introduces an inference-only scheme for agentic speech recognition where a frozen corrector uses a per-domain memory file to decide per utterance whether to act or abstain, reducing word error rate across multiple domains without weight updates.

Kernel Forge is an open-source agent harness that uses LLMs and Monte Carlo Tree Search to automatically generate and optimize CUDA kernels for any unmodified PyTorch model, achieving up to 2.83× speedup on softmax in Gemma 4 E2B.

Ethan Mollick's updated guide on which AI to use now emphasizes agentic systems over simple chat, comparing ChatGPT Work, Cowork, Codex, and Code modes, while noting Gemini's absence from the list.

A practitioner details the real-world challenges of audit, compliance, and governance for autonomous AI agents, including identity, approvals, logging, and accountability, while asking the community for solutions.

Cursor's research shows that multi-model agentic systems using a frontier model as planner and a cheaper model as worker can reduce token costs by 15x, enabling more efficient complex agents.

A survey of 239 papers analyzing how AI agents self-improve by updating the model itself or the scaffold (prompts, memory, tools).

HealthClaw is an open-source agent architecture for longitudinal personal health management that uses self-evolving memory to improve support over repeated encounters, achieving higher accuracy and privacy compared to baselines across biomedical tasks.