AgentBench AI Agent
by THUDMA Comprehensive Benchmark to Evaluate LLMs as Agents (ICLR'24)
AgentBench is an open-source ai agent by THUDM, written mainly in Python, with 3,804 GitHub stars and the Apache-2.0 license. Its code was last updated Feb 8, 2026.
Scan summary: scanned Nov 5, 2026 at commit d1e4a10. Seen able to run dynamic code, connect to the internet. Scan policy
What AgentBench can touch
- Our checks did not find this. The tool may still do it.Run shell commandsNot seen in this scan
- Seen in this scan, either in the code or while it ran.Run dynamic codeFound in the code by our scan
- Our checks did not find this. The tool may still do it.Read or write filesNot seen in this scan
- Our checks did not find this. The tool may still do it.Read environment and secretsNot seen in this scan
- Our checks did not find this. The tool may still do it.Control a web browserNot seen in this scan
- Our checks did not find this. The tool may still do it.Connect to databasesNot seen in this scan
- Our checks did not find this. The tool may still do it.Read or send messagesNot seen in this scan
- Our checks did not find this. The tool may still do it.Touch money or tradingNot seen in this scan
- Seen in this scan, either in the code or while it ran.Connect to the internetFound in the code by our scan
"Not seen" means our checks did not find it, and the tool may still do it. Covers commit d1e4a10, Nov 5, 2026. See the scan policy.
Before you install AgentBench
AgenticMart scanned AgentBench on Nov 5, 2026 with automated checks for hidden code, known-malicious packages, leaked secrets and vulnerable dependencies. The result covers one commit under our scan policy, so review what access you give it and check any advisories shown here.
What it can access
- An AI agent decides its own next steps and calls whatever tools you give it, so its risk depends on the access you grant.
Before you install
- Pin the version you reviewed instead of always installing the latest release.
- Read the tool list or scripts before the first run, and keep approval prompts on for actions that write, send or delete.
- Maintained by THUDM on GitHub, 3,804 stars and 293 forks. Popularity is not proof of safety.
- Last code change: Feb 8, 2026.
- Open-source license: Apache-2.0.
AgentBench is an open-source ai agent by THUDM, written mainly in Python. A Comprehensive Benchmark to Evaluate LLMs as Agents (ICLR'24)
At a glance
- Stars
- 3.8k
- Forks
- 293
- Language
- Python
- License
- Apache-2.0
- Created
- Jul 2023
- Last update
- Feb 8, 2026
- Type
- AI Agent
- Price
- Free
Topics
Questions about AgentBench
What is AgentBench?
A Comprehensive Benchmark to Evaluate LLMs as Agents (ICLR'24)
Who maintains AgentBench?
AgentBench is maintained by THUDM on GitHub (THUDM/AgentBench). AgenticMart lists it from public repository data.
Is AgentBench safe to use?
AgenticMart scanned AgentBench on Nov 5, 2026 with automated checks for hidden code, known-malicious packages, leaked secrets and vulnerable dependencies. The result covers one commit under our scan policy, so review what access you give it and check any advisories shown here.
What license does AgentBench use?
It is published under the Apache-2.0 license.
AgentBench alternatives
View more →Similar ai agents in Other, ranked by GitHub stars. Compare licence, recent activity and scan status before you choose.
| Name | Stars | Licence | Last updated | Scan |
|---|---|---|---|---|
| AgentBench (this page) THUDM | 3.8k | Apache-2.0 | Feb 8, 2026 | Scanned, no flagged issues |
| Hermes Desktop fathah | 15k | MIT | Oct 13, 2026 | Scanned, no flagged issues |
| CubeSandbox TencentCloud | 13k | None found | Nov 1, 2026 | Not scanned yet |
| Openchamber openchamber | 12k | MIT | Nov 1, 2026 | Scanned, no flagged issues |
| Open Swe langchain-ai | 11k | MIT | Nov 1, 2026 | Scanned, no flagged issues |
| Moltworker cloudflare | 9.9k | Apache-2.0 | May 9, 2026 | Scanned, no flagged issues |