# The Dual-State Architecture for Reliable LLM Agents

arXiv: https://arxiv.org/abs/2512.20660 · HTML: https://mlsanity.com/paper/2512.20660/

## Key facts

- Cited by 3 papers in the map (data through 14 Oct 2026).
- Placed in Software engineering coding agents, one of 16 directions in LLM agents and tool use.
- Builds on 5 of its 18 references that are in the map.
- First posted on arXiv 18 Dec 2025.

## Placed in

- LLM agents and tool use / [Software engineering coding agents](https://mlsanity.com/agents/software-engineering-coding-agents/)

## Builds on (most cited first, 5 of 5)

- [Chain-of-Thought Prompting Elicits Reasoning in Large Language Models](https://mlsanity.com/paper/2201.11903/) (2022)
- [ReAct: Synergizing Reasoning and Acting in Language Models](https://mlsanity.com/paper/2210.03629/) (2022)
- [SWE-bench: Can Language Models Resolve Real-World GitHub Issues?](https://mlsanity.com/paper/2310.06770/) (2023)
- [LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks](https://mlsanity.com/paper/2402.01817/) (2024)
- [On the Planning Abilities of Large Language Models (A Critical Investigation with a Proposed Benchmark)](https://mlsanity.com/paper/2302.06706/) (2023)

## Built on by (most cited first, 3 of 3)

- [Agentic Problem Frames: A Systematic Approach to Engineering Reliable Domain Agents](https://mlsanity.com/paper/2602.19065/) (2026)
- [Terminal Agents: A Survey of AI Agents in Command-Line Environments](https://mlsanity.com/paper/2608.20485/) (2026)
- [CAX-Agent: A Lightweight Agent Harness for Reliable APDL Automation](https://mlsanity.com/paper/2605.15218/) (2026)

## Abstract (arXiv)

Large Language Models deployed as code generation agents exhibit stochastic behavior incompatible with the deterministic guarantees required by software engineering. We formalize the Dual-State Action Pair (DSAP), an execution primitive that couples stochastic generation with deterministic post-condition verification. Guard functions act as sensing actions that project opaque LLM outputs onto observable workflow state, enabling a dual-state decomposition: finite, deterministic S_workflow paired with infinite, stochastic S_env. We prove that for epsilon-capable generators, failure probability P(fail) <= (1-epsilon)^R_max -> 0. To prevent naive O(R^K) retry explosion across multi-step workflows, we introduce a three-level recovery hierarchy: context refinement (retry within step), informed backtracking (stagnation detection with cascade invalidation and context injection to upstream steps), and human escalation. Experimental validation across 13 LLMs (1.3B-15B parameters) on three diagnostic probes demonstrates reliability gains of up to 66 percentage points at 1.2-2.1x baseline cost. Recovery mechanism evaluation on 99 SWE-Bench Pro instance-arm pairs (Qwen3-Coder-Next) demonstrates 100% context injection effectiveness (upstream output changed in all 71 escalation events) with step-specific recovery asymmetry -- 37.5% for test generation vs. 0% for patch generation -- and 0% end-to-end patch production, establishing the boundary between execution architecture and plan synthesis: execution recovery is necessary but not sufficient for autonomous software engineering.