Wink: Recovering from Misbehaviors in Coding Agents

Rahul Nanda; Chandra Maddila; Smriti Jha; Euna Mehnaz Khan; Matteo Paltenghi; Satish Chandra

cs.SEFeb 19, 2026

Wink: Recovering from Misbehaviors in Coding Agents

Authors: Rahul Nanda, Chandra Maddila, Smriti Jha, Euna Mehnaz Khan, Matteo Paltenghi, Satish Chandra

Organizations: Meta Platforms, Inc. New York, NY, USA · Meta Platforms, Inc. Bellevue, WA, USA · Meta Platforms, Inc. Menlo Park, CA, USA

Abstract

Autonomous coding agents, powered by large language models (LLMs), are increasingly being adopted in the software industry to automate complex engineering tasks. However, these agents are prone to a wide range of misbehaviors, such as deviating from the user's instructions, getting stuck in repetitive loops, or failing to use tools correctly. These failures disrupt the development workflow and often require resource-intensive manual intervention. In this paper, we present a system for automatically recovering from agentic misbehaviors at scale. We first introduce a taxonomy of misbehaviors grounded in an analysis of production traffic, identifying three primary categories: Specification Drift, Reasoning Problems, and Tool Call Failures, which we find occur in about 30% of all agent trajectories. To address these issues, we developed a lightweight, asynchronous self-intervention system named Wink. Wink observes agent trajectories and provides targeted course-correction guidance to nudge the agent back to a productive path. We evaluated our system on over 10,000 real world agent trajectories and found that it successfully resolves 90% of the misbehaviors that require a single intervention. Furthermore, a live A/B test in our production environment demonstrated that our system leads to a statistically significant reduction in Tool Call Failures, Tokens per Session and Engineer Interventions per Session. We present our experience designing and deploying this system, offering insights into the challenges of building resilient agentic systems at scale.

Figures & tables

Figure 1. Self-intervention mechanism addressing specification drift. The agent deviates from user instructions, the intervention agent detects this drift, redirects the agent to follow the explicit instructions, and the agent acknowledges and executes the correct tool invocation, successfully completing the task.

Misbehavior Category	Trajectories Detected	Prevalence
Loops	2232	5.21%
DNF	6827	15.95%
UC	2833	6.62%
Tool Call Failure	6001	14.02%
Total Misbehavior Categories	12,499	29.2%

Table 1. Prevalence of detected misbehavior categories in 42,807k trajectories sampled from production traffic over a period of five weeks.

Figure 2. Overview of the misbehavior prevalence metrics computation: Taxonomy creation and trajectory classification

Misbehavior Category	Prevalence in Sonnet 4.5 (%)	Prevalence in Opus 4.5 (%)
Did Not Follow Instructions	18.69	13.19
Unrequested Changes	6.55	5.2
Infinite loops	2.18	3.59
Tool Call Failure	10.99	11.08

Table 2. Prevalence of Misbehaviors in Sonnet 4.5 and Opus 4.5

Figure 3. The Self-Intervention system architecture. The observer runs asynchronously to prevent latency regressions in the main SWE agent loop, injecting guidance via system-reminders only when results are available.

Figure 4. Self-Intervention mechanism addressing agent infinite loops misbehavior. The diagram illustrates recovery (A) and non-recovery (B) trajectories following an intervention. Recovery occurs if the specific misbehavior (e.g., redundant calls to read_file(php_syntax.md) ) does not recur. Conversely, non-recovery is defined by the reoccurrence of the same pattern of actions constituting that misbehavior. Different instances of agent misbehavior (C) can arise later, which would trigger new, separate intervention steps.

Misbehavior Category	Sample Size	Agent Recovered	Agent Not Recovered	% Agent Recovery Rate
Reasoning Problems	963	908	55	94.29
Tool Call Failure	2604	2386	218	91.63
Specification Drift	1627	1429	198	87.83
Overall	5194	4723	471	90.93

Table 3. Recovery rate for single-intervention conversations, as determined by the LLM judge. This estimate is conservative: if the evidence is insufficient to verify recovery, the judge labels the post-intervention trajectory as “not recovered.”

Misbehavior Category	Sample Size	Agent Recovered	Agent Not Recovered	% Agent Recovery Rate
Reasoning Problems	2414	1781	633	73.78
Tool Call Failure	1531	1346	185	87.92
Specification Drift	1415	1111	304	78.52
Overall	5360	4238	1122	79.07

Table 4. Recovery rate for conversations with multiple-interventions, as determined by the LLM judge. This estimate is conservative: if the evidence is insufficient to verify recovery, the judge labels the post-intervention trajectory as “not recovered.”

Metric	$\Delta$ %
Tool Call Failures	$-4.2$ % ∗∗
Tokens per Session	$-5.3$ % ∗∗
Engineer Interventions per Session	$-4.2$ % ∗
Agent Execution Time per Session	$-4.3$ %

Table 5. Results from the live A/B test

Outcome	Reason	%
Recovered	Recover from infinite loops	39
	Remind original task with updated plan	26
	Ask agent to not overdo	17
	Provide right tool arguments	17
Not Recovered	Agent ignored the course-correction instructions	37
	Premature termination	22

Table 6. Qualitative coding categories by recovery outcome

Figure 5. Self-intervention mechanism showing a tool call failure followed by intervention and recovery.

LLM Agent Reliability

Large language model (LLM) agent reliability concerns dependable, safe, and verifiable behavior across tools, multi-step tasks, and changing environments.

See all 832 papers

SWE-MISBEHAVIOR: Benchmarking and Mitigating Boundary Violations in Software Coding Agents
Oct 13, 2026Do Xuan Trong, Duc Anh Nguyen, Do Xuan Long +12AI Coding Agents AI Agent Safety
National University of Singapore · Vietnam National University, Hanoi · University of Virginia, USA +4
Failure as a Process: An Anatomy of CLI Coding Agent Trajectories
Jul 10, 2026Xiangxin Zhao, Han Li, Shuaiting Li +4Agent Failure Analysis Coding Agents
University College London, London, United Kingdom · Nanjing University, Nanjing, China
Learning When and How to Intervene: A Hindsight-Distilled Sentinel for Coding Agents
Sep 30, 2026Jiangrui Zhao, Chenglong Li, Meng Zhang +1Coding Agents Runtime Enforcement for AI Agents
Beijing University of Posts and Telecommunications · Beijing University of Technology · Meta

Oct 13, 2026J/K move · Enter open · S save

SWE-MISBEHAVIOR: Benchmarking and Mitigating Boundary Violations in Software Coding Agents

Do Xuan Trong, Duc Anh Nguyen, Do Xuan Long, Ngoc-Hai Nguyen+11

National University of Singapore · Vietnam National University, Hanoi · University of Virginia, USA+4

AI Coding Agents AI Agent Safety AI Agent Safety Benchmarks AI Agent Security Benchmarks

Open paper AskPDF arXiv

We present SWE-MISBEHAVIOR, the first taxonomy and evaluation framework for diagnosing boundary-violating misbehaviors in coding agents. It contains 7.4K human-in-the-loop synthesized samples spanning three layers of agent action space: (i) Code Generation, (ii) File Management, and (iii) System Interaction. Across these categories, SWE-MISBEHAVIOR covers 11 distinct misbehaviors with a fully automated evaluation protocol, including coding-action violations, Create/Read/Update/Delete (CRUD) file misbehaviors, system integrity and abuse violations, and task-goal misalignment. Experiments on state-of-the-art open-source coding agents (e.g., OpenHands, smolagents) reveal substantial misaligned behaviors, including dependency misinterpretation (up to 98.5% of cases), unintended modification of user inputs (50.5%), and silent compromise of environment integrity (53%). Motivated by these findings, we further propose CODE-KIT, a lightweight adaptive framework that combines misbehavior-aware guidelines with novel tools to help agents avoid such failures. CODE-KIT reduces misbehaviors by 52% for smolagents and 34% for OpenHands, while simultaneously improving task completion performance.

Abstract

LLM Agent Reliability

Explore similar work

SWE-MISBEHAVIOR: Benchmarking and Mitigating Boundary Violations in Software Coding Agents

Failure as a Process: An Anatomy of CLI Coding Agent Trajectories

Learning When and How to Intervene: A Hindsight-Distilled Sentinel for Coding Agents