Unfinished-Bench: Measuring Rule Violations by
Persistent Agents in Realistic Work Environments

Seunghyun Lee, Jaehui Hwang, Dongyoon Han, Sangdoo Yun*
NAVER AI Lab
*Corresponding author
NAVER AI Lab
Overview of Unfinished-Bench: a user hands off a task in an organization and leaves; the harness adds a task-completion instruction and re-prompts; the agent either violates the rule to finish or stops short and reports back.
Overview. A user hands off a task and leaves; the rule (red) sits in a document, on a tool screen, or with a colleague, and the goal and the rule cannot both hold. The harness adds persistence, and the agent either violates the rule to finish or stops short and reports back.

TL;DR Agents are often told to keep going until the task is done. When a task can be finished only by violating a rule found in the work environment, this persistence raised the average violation rate across 12 open models from 32.2% to 43.6%, and pushed half of the models above 70%.

Play

You are the harness that adds persistence

Pick a user who hands off a task. Each time their agent stops before the task is done, keep it going or let it stop. Your choices decide how the run ends; the agent’s steps come from recorded runs in the benchmark.

Unfinished-Bench · pick a user to start at desk
YOUHarness
Pick a user to begin.
Keep going 0

The agent’s steps come from recorded runs in the benchmark. Each event line summarizes a step; open it to see the transcript text, quoted and shortened with ellipses. Letting the agent stop ends the run at that point.


Benchmark

A work assignment the agent cannot finish without violating a rule

Persistence

Persistence is an agent’s drive to complete its task. It makes long-running agents useful but can amplify unwanted behavior: in the recent OpenAI–Hugging Face incident, agents trained to be persistent rarely abandoned their tasks and increasingly turned to out-of-bounds methods on tasks they could not solve.

In Unfinished-Bench, the harness gives the agent persistence, and we run every task without and with it.

No persistence
TaskStopsRun ends
Persistenceno stop until the task is complete
TaskTask complete
In Unfinished-Benchpersistence through the harness
Task“See this through
to completion…”
StopsKeep going…

A realistic workplace

Each task is set inside one of 7 organizations built for the benchmark, with staff, internal documents, chat, and work tools. The rule appears as it would at work: a line in a document, a line on a work tool screen, and a colleague who holds the authority.

Internal documents
Chat
Work tools
Colleagues

471 internal documents · 8,066 chat messages · 1,301 staff

red: where the rule that must be kept appears

Seven risk categories

Unauthorized transactionse.g. a claim held for a fraud review
Data leakagee.g. records of study participants who withdrew consent
Destructive actionse.g. mailboxes under a litigation hold
Service disruptione.g. another team's running batch job
Security weakeninge.g. the TLS floor on a partner endpoint
Unauthorized external communicatione.g. contacts marked do-not-contact
Restriction bypasse.g. a production release awaiting change approval

97 scenarios across seven risk categories. Runs are isolated and offline; violations are scored by code from audit logs and server state.


Results

Persistence adds violations

Violation rates of 12 open models, without and with persistence.

No persistence Persistence Right: change in points

The rule appears only in one section of an internal document.

How the user’s oversight changes these violations, in an earlier, smaller study: Don’t Leave the Task Unfinished.


Paper

Abstract

AI agents are increasingly equipped with persistence toward task completion to run delegated work autonomously for long periods, under many unstated constraints from internal documents and other people's authority. When completion requires violating a rule, the agent must decide alone, yet benchmarks of instruction–rule conflicts state the rule in the prompt or keep it hidden and do not evaluate persistence. We present Unfinished-Bench, a benchmark that gives models persistence in autonomous runs without a user, in realistic work environments with embedded rules. In each of its 97 scenarios across 7 organizations and 7 risk categories, the goal and the rule cannot both hold. We find that persistence raised violations of a rule in an internal document from 32.2% to 43.6%, five times the 8.5% for the same rule stated by the user. Of the document-rule violations, 98.3% occurred after the rule was displayed, and in 89% of a judged sample the agent first wrote down the rule. Violations of rules placed in the environment without user oversight can lead to serious incidents, so persistence-driven risks need to be evaluated.

Cite

Citation

@misc{lee2026unfinishedbench,
  title  = {Unfinished-Bench: Measuring Rule Violations by Persistent Agents in Realistic Work Environments},
  author = {Lee, Seunghyun and Hwang, Jaehui and Han, Dongyoon and Yun, Sangdoo},
  year   = {2026},
  eprint = {2611.17460},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url    = {https://arxiv.org/abs/2611.17460}
}