Benchmarks / WorkArena
WorkArena
Built by ServiceNow · Alexandre Drouin, Maxime Gasse, Massimo Caccia, Issam H. Laradji, et al. · released 12 Mar 2024
Browser tasks on a live ServiceNow enterprise instance, the software many companies use for IT and HR requests. WorkArena-L1 has 33 task types (19,912 instances) that cover the main parts of the ServiceNow interface: forms, lists, knowledge base search, service catalog orders, and dashboards.
Frontier
63.3%
WorkArena-L1 success rate
GenericAgent-Claude-4-Sonnet · harness: BrowserGym GenericAgent
7 Aug 2025 · Source: ServiceNow (benchmark maintainers)
Standard error 2.7. Best entry that follows the standard evaluation protocol (15 steps).
The best result is at 60–90% of the ceiling, or at or above human parity.
What it measures
Whether a web agent completes routine enterprise-software actions from a text instruction. Each task samples a fresh configuration, and a checker validates the result in the ServiceNow instance. WorkArena++ (L2 and L3, 682 tasks) composes these atomic actions into longer workflows that need planning and reasoning. The leaderboard also lists L2 and L3 results; this entry scores L1.
WorkArena-L1 success rate: Share of sampled WorkArena-L1 task instances completed, as reported on the BrowserGym leaderboard with standard error. Higher is better.
Status compares the frontier with a ceiling of 100%.
Facts
- Grain
- Task level: bounded tasks with a clear spec
- Environment
- browser, live-system
- Grading
- state-check
- Tasks
- 33
- Human reference
- None for L1. In the WorkArena++ paper, 15 volunteers (about 75% of them current or former ServiceNow staff) solved 93.9% of a 98-instance subset of L2 and L3 tasks; a GPT-4o agent solved 2.1% of the same subset.
- Contamination
- Task templates are public, but each run samples new instances. The maintainers have not published a contamination analysis.
- Reuse
- Apache 2.0 for the benchmark code. Running tasks needs access to a ServiceNow developer instance. (open-apache)
Limits to keep in mind
- Leaderboard entries are self-submitted result files. The two highest entries raised the step budget from 15 to 30, and their files mark them as not following the evaluation protocol. We show the best entry that follows it (63.3%). Source
- L1 tasks are short atomic actions, so L1 overstates how close agents are to full workflows. On the same board the best WorkArena++ L2 score is 69.4% (GenericAgent-GPT-5), and every L3 entry that follows the protocol is below 1%. Source
- Few agents have been submitted (about 20 folders), and several have no WorkArena result at all. Source
- All tasks run on one vendor's platform, so results say little about other enterprise software. Source
Recorded results
A selection that shows the frontier over time. The full leaderboard is at the source.
| System | WorkArena-L1 success rate | Date | Source |
|---|---|---|---|
| IpaziaHPA-Gemini-3-flash-preview harness: IpaziaHPA | 90.3% | 7 Mar 2026 | Ipazia S.p.a. (submitted to the ServiceNow BrowserGym leaderboard) · third-party |
| GenericAgent-Claude-4-Sonnet · frontier harness: BrowserGym GenericAgent | 63.3% | 7 Aug 2025 | ServiceNow · primary |
| GenericAgent-GPT-5 harness: BrowserGym GenericAgent | 79.1% | 7 Aug 2025 | ServiceNow · primary |
| GenericAgent-GPT-4o harness: BrowserGym GenericAgent | 45.5% | 23 Oct 2024 | ServiceNow · primary |
Timeline
- 12 Mar 2024 — ServiceNow Research releases WorkArena and BrowserGym. Source
Where it sits in the atlas
Work ladder: Direct evidence for Office and administrative support. On the work ladder it counts as 63.3% of tasks completed. How the ladder works
Work it measures (O*NET work activities): Perform administrative or clerical activities; Process digital or online data; Gather information from physical or electronic sources.
Office and administrative supportDevOps, SRE, and IT operations (partial)
Go to the source
- Website github.com
- Paper arxiv.org
- Full leaderboard huggingface.co
- Code github.com
Last checked 10 Oct 2026 against 10 primary sources. See an error? Tell us.
How to cite
Credit the original work first: WorkArena by ServiceNow (https://arxiv.org/abs/2403.07718).
Then, if you used this page:
Can Agents Work. "WorkArena: frontier results and sources." https://canagentswork.com/benchmarks/workarena/ (accessed 2026-10-11). CC BY 4.0.@misc{caw-workarena,
title = {{WorkArena: frontier results and sources}},
author = {{Can Agents Work}},
year = {2026},
howpublished = {\url{https://canagentswork.com/benchmarks/workarena/}},
note = {Accessed 2026-10-11. CC BY 4.0}
}