Source-linked AI summary
Agentic Transaction: Towards ACID-Compliant Agent Systems
Zhaoyan Sun, Xiaoxiao Wang, Guoliang Li
TL;DR
Long-horizon agent workflows need database-like guarantees for reliable execution, consistency, concurrency, and durable state. This paper introduces an ACID-compliant agent framework and data agent, which outperforms Claude Code by 10.6% in overall score on KramaBench.
Problem
Long-horizon agent workflows involve dynamic, multi-step actions and non-transactional resources, creating a gap in reliable management of partial effects and semantic state.
Method
The paper defines agentic transactions and implements ACID-inspired guarantees through transactional execution cycles, confidence-based validation, semantic isolation, and transaction-aware state management.
Results
10.6% overall-score improvement over Claude Code was achieved by ACID-Agent with the Qwen3.5-197B-A17B backbone on KramaBench.
Takeaways & Limitations
The framework provides a principled foundation for designing more reliable agent systems with transactional execution semantics.
Takeaways & Limitations
General-purpose ACID-compliant agents remain an open problem, including scalable skill ecosystems, stable reasoning, and safe multi-agent coordination.
Abstract
from arXiv · showhide
Large language model (LLM) agents are evolving from conversational assistants into autonomous systems that execute long-horizon tasks through reasoning, tool use, code generation, and workspace manipulation. As agents increasingly operate over persistent environments and multi-step workflows, they face challenges analogous to those addressed by transactional database systems: reliable execution, consistent outcomes, safe concurrency, and durable state management. We introduce the concept of an agentic transaction and propose an ACID-compliant agent system framework that reinterprets the classical ACID properties for agent execution through four semantic guarantees: Semantic Atomicity, Semantic Consistency, Semantic Isolation, and Semantic Durability. Together, these properties provide a principled foundation for building reliable agent systems despite model uncertainty and dynamic execution environments. To instantiate this framework, we develop an ACID-compliant data agent that realizes these guarantees through transactional exploration-execution-validation cycles, transactional skill hubs, confidence divergence-based validation, semantic dependency-aware isolation, and transaction-aware semantic state management. Experimental results on widely used benchmarks show that our system achieves a 10.6% improvement over state-of-the-art agents, including Claude Code. This work opens a broader research agenda on extending transactional principles and system architectures toward building trustworthy, scalable, and self-evolving AI agent systems.
1 INTRODUCTION
The paper frames long-horizon LLM agents as systems that need transactional guarantees because they coordinate iterative reasoning, code execution, tool use, and workspace changes over persistent environments. It defines agent transactions as dependency-aware semantic execution units and proposes ACID-style mechanisms for validation, isolation, and durable state management despite nondeterminism and non-transactional effects.
- Motivation: LLM use is shifting from single-round conversation toward long-horizon production tasks spanning repository workspaces, iterative reasoning, code execution, and feedback-driven refinement.These capabilities motivate treating agent execution as a systems problem rather than only a conversational one.
- Agentic Transactions: An agent transaction groups dependency-aware model invocations, tool calls, document mutations, and external actions into one semantic execution unit.Effects become visible only after required operations and validations succeed; otherwise, recoverable effects are rolled back or compensated.
- Challenges: Agent workflows complicate transaction design because they are long-running, dynamically generated, and often interact with non-transactional external resources.Partial execution can therefore leave inconsistent state, motivating semantics that constrain committed effects rather than deterministic execution traces.
- Semantic Validation: ACID-Agent validates execution using execution errors, decision and code confidence divergences, and LLM-based reflection grounded in exploration evidence.It triggers refinement when confidence falls below a threshold, linking validation to observed reliability signals.
- Semantic Durability: ACID-Agent preserves durability through an append-only workspace that records committed transaction states and updates memory only with validated execution units.Failed attempts are discarded despite distributed state across conversations, model outputs, tools, files, databases, and external services.
2 AGENTIC TRANSACTION
An agentic transaction is a bounded sequence of LLM-driven interactions with an execution environment that accomplishes a task. The proposed ACID-compliant data agent provides reliability guarantees through validated transaction units, semantic validation, dependency-aware isolation, and durable state management.
- Agentic Transaction: An agentic transaction is a bounded, finite sequence of LLM-driven interactions between an agent and its environment undertaken to accomplish a task.Each step includes an LLM context, an agent action invoking tools or skills, and execution feedback.
- Agentic Transaction: Exploration-execution-validation cycles form semantic transaction units whose effects are propagated only after validation, while failed units are discarded or recovered.This realizes commit-or-retry behavior without affecting committed state.
- ACID-Compliant Data Agent: The ACID-compliant data agent targets reliability across data exploration, execution, and transactional state evolution.Its design is based on agentic transaction principles and addresses data science workflows over heterogeneous data and data systems.
- Semantic Atomicity: Semantic atomicity combines an offline skill hub with online staged execution that uses validation gates and commits only validated effects.Tools and skills are treated as first-class transactional objects with lifecycle interfaces.
- Semantic Consistency: Confidence-based validation detects execution deviations using errors, confidence divergence, and reflection feedback, triggering retries when predefined thresholds are exceeded.Confidence is quantified from token-level log probabilities and compared across contexts.
- Semantic Isolation and Durability: Semantic isolation addresses dependencies among agents, contexts, workspaces, and operations through agent-level and operation-level isolation, while semantic durability preserves state with tracing and recovery.Durability combines transaction-aware semantic state management with persistent execution tracing and recovery.
3 EXPERIMENTS
Experiments validate ACID-Agent on the multi-step KramaBench benchmark, showing higher overall performance than Claude Code and benefits from transactional isolation and execution design.
- Experimental Setup: KramaBench contains 104 natural-language tasks over 1,700 real-world data files from 24 sources across six domains.Tasks require reasoning over heterogeneous datasets and multi-step workflows.
- Overall Evaluation: ACID-Agent is evaluated on overall performance and execution consistency, with consistency measured by average per-task variance across three independent runs.The evaluation compares task scores and run-to-run consistency across LLM backbones.
- Overall Evaluation: 10.6% overall-score improvement: ACID-Agent powered by Qwen3.5-197B-A17B outperforms Claude Code.The passage reports higher overall scores across different LLM backbones, with improvements observed in most domains.
- Ablation Study: 11.7% score reduction: removing failed-step isolation allows intermediate failures to update workspace and context memory, contaminating subsequent execution.This ablation demonstrates the importance of isolating failed steps.
- Ablation Study: ACID-Agent outperforms majority-voting Claude Code across three runs while using fewer tokens, indicating its improvement does not come from increased token usage.The passage reports this as the effect-of-more-tokens ablation.
4 OPEN PROBLEMS
General-purpose ACID-compliant agentic systems remain an open research challenge beyond data agents. Key questions concern scalable and safe skill ecosystems, reliable reasoning and stable execution, and supporting system-level guarantees.
- Scope: General-purpose agentic systems must extend ACID compliance beyond data agents.The passage frames this extension as an open research challenge.
- Atomicity: Atomicity requires scalable skill ecosystems with executable semantics, quality assurance, and safety guarantees for reliable skill composition and evolution.These requirements are posed as open questions for building reliable skill ecosystems.
- Consistency: Consistency requires reliable reasoning and stable execution across runs through new architectures, benchmarks, model alignment techniques, and execution harnesses.The passage identifies these mechanisms as directions for ensuring consistency.