PARC: An Autonomous Self-Reflective Coding Agent for Robust Execution of Long-Horizon Tasks (2512.03549v1)

Abstract: We introduce PARC, a coding agent for the autonomous and robust execution of long-horizon computational tasks. PARC is built on a hierarchical multi-agent architecture incorporating task planning, execution, and a mechanism that evaluates its own actions and their outcomes from an independent context and provides feedback, namely self-assessment and self-feedback. This design enables PARC to detect and correct high-level strategic errors and sustain progress without human intervention. We evaluate PARC across computational science and data science tasks. In materials science, it autonomously reproduces key results from studies on lithium-ion conduction and alloy segregation. In particular, it coordinates dozens of parallel simulation tasks, each requiring roughly 43 hours of computation, managing orchestration, monitoring, and error correction end-to-end. In Kaggle-based experiments, starting from minimal natural-language instructions, PARC conducts data analysis and implements search strategies, producing solutions competitive with human-engineered baselines. These results highlight the potential of integrating a hierarchical multi-agent system with self-assessment and self-feedback to enable AI systems capable of independent, large-scale scientific and analytical work.

• The paper introduces PARC, a hierarchical coding agent system that decomposes user objectives into sub-tasks and employs self-reflection for error detection and correction.
• It details a dual-component architecture with a planner coordinating multiple worker agents to execute tasks while managing context modularity and preventing cross-task contamination.
• Empirical studies in simulations and data science benchmarks demonstrate PARC’s competitive accuracy and reliability, highlighting both its strengths and the remaining challenges in error coverage.

Autonomous Long-Horizon Task Execution with Self-Reflective Agents: A Technical Appraisal of PARC

Introduction

The paper "PARC: An Autonomous Self-Reflective Coding Agent for Robust Execution of Long-Horizon Tasks" (2512.03549) introduces PARC, a hierarchical multi-agent coding system engineered for the robust, autonomous execution of long-horizon computational workflows. PARC distinguishes itself in the current landscape of coding agents by integrating explicit self-assessment and self-feedback mechanisms. The purpose of this essay is to provide a detailed technical summary of the architectural design and empirical results of PARC, highlighting its contributions, performance characteristics, and broader implications for agent-based AI design and long-horizon automation.

Architectural Overview and Agent Design

PARC employs a hierarchical architecture comprising two primary components: a planner and multiple worker agents. The planner is responsible for decomposing the user’s objective and high-level instructions into a sequenced series of executable tasks. Each worker agent then operates within a dedicated context to execute a specific task, thereby circumventing issues of context window saturation and cross-task contamination encountered in monolithic agent designs. This plan-and-execute protocol ensures context modularity and scalability.

A defining property of PARC is its systematized self-reflection. Both during and after task execution, agents invoke self-assessment routines that evaluate the correctness, consistency, and validity of generated artifacts, including code, data outputs, and analytical results. This feedback loop is architected to address both local failures—such as code bugs or pipeline crashes—and strategic, approach-level errors that may arise due to flawed planning or invalid assumptions.

Figure 1: Schematic illustration of the PARC hierarchical workflow, including planner-task decomposition and worker-task execution with shared workspace and self-reflective feedback.

These architectural mechanics enable PARC to perform robust error handling, context inheritance through a structured workspace, and dynamic workflow adaptation. The agent autonomously manages dependencies, aggregates context from prior steps, and provides long-horizon stability by halting, revising, or adapting workflows as needed.

Empirical Evaluation: Computational Science Case Studies

PARC’s efficacy is empirically validated across several computational science domains, notably materials simulation and alloy modeling.

Lithium Diffusion in Solid Electrolytes

In a simulation paper of lithium-ion diffusion in $\mathrm{Li_{10}GeP_2S_{11.5}O_{0.5}}$ (LGPS), the agent reproduces the workflow of [Sawada et al., 2024], including random structural search, molecular dynamics at multiple temperatures, and activation energy extraction through Arrhenius analysis. Remarkably, PARC centrally orchestrates dozens of parallel simulations, each exceeding 40 hours of runtime, and autonomously manages the entire process from environment setup to final report generation.

Numerically, PARC attains an activation energy estimate of 0.23 eV, closely matching the reference value of 0.18 eV from the literature, with discrepancy attributed to structural randomness and finite-sample effects. The self-reflection mechanism is instrumental, dynamically identifying unreliable statistical conditions and prompting workflow recomputation under more stringent parameters.

Figure 2: Output sequence and comparative diffusion analysis on LGPS simulations—MSD evolution, Arrhenius fitting, and structural visualization.

Alloy Segregation Analysis

In modeling Cr-Ni alloys doped with light interstitials (B and N), PARC executes complex MC simulation workflows, incorporating custom transition rules, parallel job handling, and polyhedral template matching. The agent quantitatively replicates compositional and structural trends from [Dolezal et al., 2025]—FCC phase degradation with B doping and its preservation under N doping—while autonomously resolving cell representation errors and parameter flaws.

Figure 3: Task decomposition and output plots from alloy simulations, detailing phase fraction evolution and atomic segregation behavior.

Notably, execution involved 35 parallel MC runs with up to 43-hour runtime per task, managed without human oversight; this level of orchestration underscores the value of autonomous, error-correcting agent architectures.

Non-equilibrium MD Simulations and Failure Modes

Attempts to reproduce the results of [Hisama et al., 2023] on YSZ ion conduction under external electric field revealed the current boundaries of autonomy. While PARC correctly implements core simulation routines and adapts to failed NPT runs via alternative energy minimization, analysis-phase failures (incorrect MSD across periodic boundaries) escape self-correction, demonstrating incomplete coverage in assessment routines. Expert intervention rectifies and validates the simulation outputs post hoc.

Figure 4: Task workflow and output comparisons for oxygen-ion displacement and conductivity under electric fields, highlighting agent/human post-processing distinctions.

Data Science and Algorithmic Workflows: Kaggle Challenges

Polymer Property Prediction

Evaluating on the NeurIPS Open Polymer Prediction Challenge, PARC autonomously designs and executes a predictive modeling workflow from minimal instructions. It achieves an average $R^2$ of 0.669 (without external tool mordred), outperforming prior published results, and 0.781 (with mordred), surpassing human-crafted public notebook baselines. The agent not only detects and rectifies data leakage but selectively applies hyperparameter optimization to avoid score regression on certain targets.

Figure 5: Task sequence overview for polymer property prediction, encompassing feature engineering, modeling, and validation.

Polytope Permutation Puzzle

Engaging with polyhedral puzzle-solving (Rubik’s Cube analog), PARC autonomously builds an emulator, implements and tests multiple search paradigms (A*, beam search, simulated annealing), dynamically categorizes puzzles by scale, and orchestrates solution post-processing. While unable to replicate top-tier solutions without external solvers, it reduces total moves from 1,220,590 to 1,199,430—competitive with strong human-engineered baselines under equivalent constraints.

Figure 6: Task sequence decomposition for the polytope puzzle, involving iterative algorithmic refinement and resource management.

Strong Results, Contradictory Claims, and Architectural Implications

Key numerical highlights:

• In scientific simulation tasks, PARC autonomously completes workflows of 10-20 tasks (~100 execution steps, several days runtime) with domain-expert-validated accuracy.
• In data science benchmarks, autonomous model quality meets or exceeds non-trivial human baselines absent domain-specific tuning.
• Contradictory to contemporary beliefs attributing most agent limitations to LLM deficits, the paper claims that agent architecture—specifically hierarchical planning and self-feedback—is a dominant limiting factor for long-horizon reliability, separable from LLM capability saturation.

These results suggest that integration of explicit self-reflection and plan modularity can mitigate exponential decay in success probability due to error accumulation and context overflow. While PARC's error detection is not yet perfect (not all semantic or requirement-level errors are recognized), the architecture demonstrably advances the reliability envelope for automation in scientific and complex engineering domains.

Theoretical and Practical Implications; Future Prospects

The work formalizes a paradigm wherein agent-level autonomy for extended workflows is paramount. The separation of local execution from high-level reflection allows for the propagation and correction of both procedural and strategic failures. Practically, this opens avenues for unsupervised software engineering, computational science, and data analytical tasks—previously constrained by human monitoring and intervention.

Theoretically, PARC embodies a move toward integrating "System 2"-esque deliberative validation over "System 1" generative routines. It provides a blueprint for future architectures where external-tool autodiscovery, task granularity optimization, and recursive workspace inspection can extend autonomous execution even further. The scaling limitation is shifted from model capacity to architectural coherence and error propagation control.

Conclusion

PARC represents an agent design that, through hierarchical planning and autonomous self-reflection, robustly executes long-horizon tasks previously considered intractable for coding agents in fields ranging from materials simulation to data science and algorithmic engineering. Numerical results validate the claims of enhanced reliability and competitive accuracy versus strong human-engineered baselines. Key limitations remain in error coverage and external tool autodiscovery, pinpointing future directions for agent architecture refinement. The implications for automation of complex workflows are substantive and motivate ongoing research in agent-based reasoning, error correction, and cross-domain autonomy.