What's changed: agents now have mechanisms to remember and correct their own mistakes
Older coding agents often treated each step in isolation, so an error in one step could silently cascade. Recent systems explicitly build in self-reflection and error-correction loops. MOSAIC, a multi-agent framework for scientific coding, uses a student-teacher paradigm where agents self-reflect, create rationale, code, and debug—this stepwise decomposition and targeted error correction directly addresses the problem of repeating mistakes across chained subproblems [2]. Similarly, AlphaEvolve uses an evolutionary approach: it continuously receives feedback from evaluators and iteratively improves the algorithm by making direct code changes, which means mistakes are not just fixed but also prevented from recurring in future iterations [3].
The key shift is from one-shot generation to iterative refinement. Aviary, a training environment for language agents, formalizes scientific tasks as cycles of actions and observations, and shows that with online training and scaling inference-time compute, even open-source models can match or exceed frontier LLMs and human experts on multi-step tasks like DNA cloning and protein engineering [4]. This suggests that the ability to avoid repeated mistakes is not just a matter of model size but of building a feedback loop into the agent's workflow.
Evidence that these mechanisms work: from scheduling to matrix multiplication
The most striking successes come from evolutionary and multi-agent designs. AlphaEvolve, applied to real computational problems, found a more efficient scheduling algorithm for data centers, simplified hardware circuit design, and even discovered a new algorithm for multiplying 4x4 complex matrices using 48 scalar multiplications—the first improvement over Strassen's algorithm in 56 years [3]. This shows that iterative feedback can not only avoid past mistakes but also push beyond human-designed solutions.
MOSAIC outperforms existing approaches on scientific coding benchmarks in accuracy, robustness, and interpretability, specifically because its design mitigates LLM hallucinations when solving complex tasks with chained subproblems [2]. HyperAgent, a generalist multi-agent system that mimics human developer workflows, also sets new benchmarks on GitHub issue resolution (SWE-Bench) and repository-level code generation, suggesting that a structured multi-agent pipeline can handle long-horizon tasks without repeating errors [6].
The catch: long-horizon tasks still expose weaknesses, especially in novel or composite scenarios
Despite these advances, the evidence is mixed. DeepSWE, a benchmark of 113 original, long-horizon engineering tasks, shows that frontier agents cluster on leaderboards but separate across a wider score band when tasks are truly novel and not contaminated by pretraining data [1]. This means that even the best agents can still fail on tasks that require genuine problem-solving rather than recall.
Another limitation appears in robotic scientific experiments: vision-language-action (VLA) models fine-tuned for atomic tasks often fail when asked to compose those tasks into longer workflows, because they miss transitional steps. An LLM-based inference plugin increased average success per atomic task by 42% during inference, but this was an intervention, not a built-in capability [5]. So while agents can be taught to avoid mistakes, the ability to generalize across novel combinations remains a challenge that requires explicit design.
About These Sources
This answer is built on 6 studies (all preprints) — published from 2024 to 2026, 6 from 2024 or later — selected as the most relevant from 6 studies that passed quality screening, drawn from 36 papers retrieved from a database of over 500 million.
Sources used in this answer
DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks
DeepSWE benchmark of 113 original, long-horizon tasks shows frontier agents separate across a wider score band than on contaminated benchmarks, indicating that avoiding mistakes on truly novel tasks is still hard.
MOSAIC: Multi-agent Orchestration for Task-Intelligent Scientific Coding
MOSAIC, a multi-agent framework with self-reflection and student-teacher debugging, outperforms existing approaches on scientific coding benchmarks, improving accuracy, robustness, and interpretability.
AlphaEvolve: A coding agent for scientific and algorithmic discovery
AlphaEvolve, an evolutionary coding agent, iteratively improves code via evaluator feedback and achieved breakthroughs like a 48-scalar-multiplication algorithm for 4x4 complex matrices, the first improvement over Strassen's in 56 years.
Aviary: training language agents on challenging scientific tasks
Aviary, a training environment for language agents, shows that with online training and scaling inference-time compute, open-source models can match or exceed frontier LLMs and human experts on multi-step scientific tasks at up to 100x lower cost.
Sci-VLA: Agentic VLA Inference Plugin for Long-Horizon Tasks in Scientific Experiments
Sci-VLA, an LLM-based inference plugin, increases average success per atomic task by 42% in long-horizon robotic scientific experiments by generating transitional actions, but requires external intervention.
HyperAgent: Generalist Software Engineering Agents to Solve Coding Tasks at Scale
HyperAgent, a generalist multi-agent system with planner, navigator, editor, and executor agents, sets new benchmarks on SWE-Bench, RepoExec, and Defects4J, demonstrating effective long-horizon task handling.