Why checking every step beats checking only the final answer
The core insight from the research is that mistakes in long tasks often happen in intermediate steps, and if you only check the final output, you miss the root cause. A 2023 study by OpenAI researchers compared two ways to train a model: giving feedback only on the final answer (outcome supervision) versus giving feedback on each intermediate reasoning step (process supervision). The process-supervised model solved 78% of problems from a challenging math dataset, significantly outperforming the outcome-supervised model [3]. This means that to avoid repeating mistakes, you need to catch and correct errors as they happen, not just at the end.
This finding is reinforced by a 2025 system called Sherlock, which analyzes agentic workflows to identify which steps are most error-prone and attaches verifiers only where needed. In tests, Sherlock improved accuracy by 18.3% on average compared to a baseline with no verification [2]. The key is that verification is selective—not every step needs a costly check, but the ones that are likely to fail do. This approach balances reliability with efficiency, which is crucial for real-world deployment.
How rollback and reusable workflows stop errors from compounding
Another way deployment workflows avoid repeated mistakes is by rolling back to the last verified state when a step fails, rather than letting the error propagate. Sherlock does exactly this: it speculatively executes downstream tasks while verification runs in the background, and if verification fails, it rolls back to the last verified output [2]. This prevents a small error from snowballing into a bigger one. The same principle appears in SKILL.nb, a 2026 framework that stores workflows as versioned notebooks with validation gates. When a step's executable code fails due to environment drift, it falls back to natural-language guidance, preserving performance even when the underlying system changes [1].
Reusing past workflows can also help avoid mistakes, but only if the workflows are governed properly. SKILL.nb found that without lifecycle management, reused workflows often fail under new conditions. In their tests, SKILL.nb retained 91.7% of initially successful tasks across three re-executions, 15.5 percentage points higher than the next best method [1]. This shows that simply saving a workflow isn't enough—you need to monitor its performance and adapt it when the environment changes, otherwise you'll repeat the same mistakes in new contexts.
What's the catch? Verification costs time and money, and not all tasks are equal
The main trade-off is that verification adds latency and cost. Sherlock's selective approach reduces verification cost by 26% compared to a Monte Carlo search-based method, but it still adds overhead [2]. The 2023 process supervision study also noted the high cost of human feedback, which is why they used active learning to label only the most informative steps [3]. So, while step-by-step verification is powerful, it's not free—you have to decide where to spend your verification budget.
Also, the evidence comes from specific domains: math reasoning, web automation, and IT workflows. A 2021 study on deploying solar-powered UAVs for communication recovery used agent evaluation algorithms to plan paths, but it focused on physical constraints like energy and attitude errors, not on LLM reasoning [4]. And a 2023 paper on chaining AI agents in cloud platforms discussed orchestration strategies but provided no quantitative results [5]. So, the strongest evidence for avoiding repeated mistakes comes from LLM-based reasoning and web tasks, not from all possible agent applications. If you're deploying agents in a different domain, you may need to adapt these principles.
About These Sources
This answer is built on 5 studies (1 peer-reviewed, 4 preprints) — published from 2021 to 2026, 2 from 2024 or later, 1 in Q1 journals, collectively cited 71 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 34 papers retrieved from a database of over 500 million.
Sources used in this answer
SKILL.nb: Selective Formalization and Gated Execution for Durable Agent Workflows
SKILL.nb, a framework for governing reusable agent workflows, achieved 53.7% single-round success on WebArena-Verified (3.9 points above baseline), retained 91.7% of initially successful tasks across three re-executions (15.5 points above next best), and recovered 72.9% of subsequent failures with only 4.2% regressions, demonstrating that lifecycle governance and gate-conditioned execution improve reliability.
Sherlock: Reliable and Efficient Agentic Workflow Execution
Sherlock, a system that selectively attaches verifiers to error-prone nodes in agentic workflows and uses speculative execution with rollback, improved accuracy by 18.3% on average over a non-verifying baseline, reduced execution time by up to 48.7% over non-speculative execution, and cut verification cost by 26% compared to a Monte Carlo search-based method.
Let's Verify Step by Step
In a study on training LLMs for math reasoning, process supervision (feedback on each intermediate step) significantly outperformed outcome supervision (feedback on final result), with the process-supervised model solving 78% of problems from a representative subset of the MATH test set, and active learning further improved the efficacy of process supervision.
Agent Evaluation in Deployment of Multi-SUAVs for Communication Recovery
In a simulation study for deploying solar-powered UAVs for communication recovery, the authors proposed two path-planning algorithms (DCPPA and GCSPPA) to evaluate agent assignments, establishing conditions for fast convergence, and demonstrated accuracy and effectiveness across different scales, but this is not about LLM reasoning and provides no quantitative accuracy figures.
Chaining AI Agents in PaaS Architectures for Multi-Step Workflow Automation
A 2023 paper explores chaining multiple AI agents within PaaS architectures for multi-step workflow automation, presenting architectural designs and orchestration strategies with case studies in IT incident response, but it provides no quantitative results or empirical evaluation.