AIAny
Icon for item

Agentic Transaction: Towards ACID-Compliant Agent Systems

Defines "agentic transactions" and an ACID-style reliability framework for LLM agents that manage long-horizon tasks over persistent environments. Implements an ACID-compliant data agent using exploration–execution–validation cycles, confidence-divergence checks, semantic isolation, and append-only durable workspaces.

Introduction

Long-running LLM agents routinely break not because models hallucinate occasionally, but because multi-step, side-effecting workflows lack transactional boundaries, reliable validation, and durable provenance. The paper's core insight is to reinterpret database ACID properties as semantic guarantees for agentic transactions, creating a principled chassis for dependable autonomous workflows.

Key Findings
  • Recasting ACID for agents: defines Semantic Atomicity, Consistency, Isolation, and Durability so committed outcomes satisfy task pre/postconditions and evidence obligations despite nondeterministic LLM traces — so what: gives a checkable correctness target for agent systems.
  • Practical harness: an ACID-compliant data agent that structures work as exploration–execution–validation cycles, with transactional skill hubs and staged execution — so what: enables commit-or-retry semantics that reduce invalid side effects.
  • Validation and isolation mechanisms: uses confidence-divergence validation of critical decisions and semantic dependency-aware isolation (isolated envs, versioned workspaces) — so what: detects unsupported claims and prevents semantic interference across concurrent transactions.
  • Durability and audit: append-only workspace, provenance-rich execution traces, and transaction-aware memory to support recovery and long-horizon reasoning — so what: makes past executions reconstructable and auditable beyond transient model contexts.
  • Empirical takeaway: prototype shows consistent gains (reported ~10.6% overall improvement vs. a strong baseline) at the cost of extra code steps and token consumption due to exploration and retries — so what: transactional discipline buys reliability but increases resource use.
Who It's For and Tradeoffs

Great fit if you build autonomous agent runtimes or data agents that perform multi-step, side-effecting work (automation pipelines, data engineering agents, long-horizon code generation) and need principled correctness, auditability, and safe concurrency. Look elsewhere if your tasks are single-step, purely conversational, or cannot tolerate the added latency and token/code overhead from exploration, validation, and versioned workspaces. The approach relies on heuristic validation signals (confidence divergence, execution traces) and careful engineering of isolation and rollback, so it’s most useful when you can invest in transactional tooling and provenance capture.

Where It Fits

This work sits between LLM agent design and dependable systems: it borrows transactional thinking from databases and adapts it to the messy semantics of tool use, prompts, and external side effects. Expect to pair these ideas with versioned repositories, lightweight local validators, and workflow orchestration when building production-grade autonomous agents.

Information

  • Websitearxiv.org
  • OrganizationsTsinghua University
  • AuthorsZhaoyan Sun, Xiaoxiao Wang, Guoliang Li
  • Published date2026/08/14

Categories

More Items

Trains a foundation GUI agent using a closed-loop, environment-grounded data stack plus in-context multimodal demonstrations to automate long-horizon desktop workflows. Combines scalable task generation/verification, subtask-level demo guidance, and a 100-task OSWorkerBench benchmark to improve strict success and task progress.

Presents a unified black-box reinforcement learning framework to train and optimize agents running inside complex execution harnesses. Uses sandbox-parallel rollouts, a serving proxy that captures model calls and reconstructs multi-turn trajectories as prefix trees, and adapted GRPO/PPO optimizers to achieve stable, scalable RL across heterogeneous harnesses.

Evaluates and trains multimodal agents to construct interactive 3D open worlds from user queries — provides a large benchmark of assets, seed worlds, and reverse-synthesized queries plus a sandbox RL gym for tool-driven editing and rubric-based verification. Reports that frontier MLLMs perform under 60% and that RL fine-tuning improves precise 3D editing.