Designing resilient toolkits and scalable RL environments for CAMEL terminal agents
# Clone the repository
git clone https://github.com/camel-ai/seta.git
cd seta
bash setup.sh#=========================================
# Run single developer agent / workforce
#=========================================
cd evaluation/terminal_bench_run/
bash run_agent.sh \
-a <attempt,0..n> \
-n <total_attempts> \
-e <conda env name> \
-w <use_workforce> # can have a try, focus on single chat agent now.└── play-zork
└── play-zork.1-of-1.test_run # trial name
├── CAMEL_WORKDIR # not used at the moment
├── agent-logs # not used at the moment
├── commands.txt # not used at the moment
├── chatagent.log # ❗️❗️ full history of running agent including test results
├── eigent_logs.json # ⚠️ exists only when running workforce
├── panes # not used at the moment
└── sessions # session logs
├── agent.cast # not used at the moment
├── agent.log # not used at the moment
├── session_logs # ❗️❗️session logs for terminal toolkit
│ ├── blocking_commands.log # ❗️❗️all block mode commands + output
│ ├── session_run_zork_1_correct_path.log # ❗️❗️non-block mode single session command + output
│ ├── session_zork-1.log # ❗️❗️same as above session_{id}.log
│ └── session_zork_start.log # ❗️❗️same as above session_{id}.log
├── tests.cast # not used at the moment
├── tests.log # ❗️❗️test log
└── tests.log.strip # ❗️❗️test log with ansi control characters removed
cd evaluation/terminal_bench_eval/
# terminal bench 1.0
bash run_eval.sh
# terminal bench 2.0
bash run_tb2.sh
## The agent class is implemented in tbench_camel_agent.py - final results will be in `evaluation/terminal_bench_eval/run/{run_id}/results.json`
- task specific terminal session logs will be in `evaluation/terminal_bench_eval/logs/camel_logs/{task_id}/`
Everything is under training folder
Please refer to Training Setup for detailed instructions.
Note: new TerminalToolkit design document Terminal Toolkit Design
@misc{seta,
author = {Qijia Shen and Jay Rainton and Aznaur Aliev and Ahmed Awelkair and Boyuan Ma and Zhiqi (Julie) Huang and Yuzhen Mao and Wendong Fan and Philip Torr and Bernard Ghanem and Changran Hu and Urmish Thakker and Guohao Li},
title = {{SETA: Scaling Environments for Terminal Agents}},
year = {2026},
month = jan,
url = {https://github.com/camel-ai/seta},
note = {Blog: \url{https://eigent-ai.notion.site/SETA-Scaling-Environments-for-Terminal-Agents-2d2511c70ba280a9b7c0fe3e7f1b6ab8}}
}
