Simple by design. Effective over long horizons.
A simple yet effective AI agent for learning, experimentation, and real work that takes more than one model turn.
English · 简体中文
Overview · Results · Quick Start · Documentation
Simple Long Horizon Agent is small enough to understand and change, while capable enough to take on real tasks with language models and tools. It gives you a practical agent to learn from, experiment with, and adapt—without requiring a large framework first.
Long-horizon work cannot be completed in a single response. It requires sustained progress across many steps: planning, acting, checking results, and iterating until the goal is complete. Simple Long Horizon Agent is designed for that shape of work while keeping the project clear and approachable.
- Simple by design — a compact project with a short setup path and concepts you can understand by inspection.
- Effective in practice — built to use tools, interact with real environments, and complete meaningful tasks.
- Long-horizon ready — designed for sustained progress, repeated tool use, verification, and iteration.
- Easy to adapt — a practical base for classes, research ideas, benchmarks, and focused team workflows.
- Built to be evaluated — runnable examples and benchmark integrations make behavior and results inspectable.
We focus on long-horizon evaluations across software engineering, terminal work, and autonomous model post-training. Reproducible results will be published here with the exact model, agent setup, and cost.
| Benchmark | Model | Score ↑ | Baseline | Δ vs. Baseline ↑ | Cost / Task |
|---|---|---|---|---|---|
| SWE-bench Pro1 Resolved (%) |
GPT-5.4 (xHigh) | 63.20% | 59.10% | +6.94% | $7.7823 |
| Terminal-Bench 2.1 official score |
GPT-5.3-Codex (xHigh) | 77.53% | 64.70% | +19.83% | $0.5667 |
| PostTrainBench2 Weighted average |
GPT-5.5 (xHigh) | 45.88% | 43.97% | +4.34% | — |
Δ vs. Baseline is the score improvement over a baseline using the same model and task budget. It separates the value of the agent from the capability of the underlying model.
An em dash means “not published yet,” not zero.
Simple Long Horizon Agent supports Python 3.10 and newer and uses uv for its environment.
git clone https://github.com/simple-agent-lab/simple-long-horizon-agent.git
cd simple-long-horizon-agent
uv sync
bash runs/demos/run_bash_agent_demo.shThe default demo is deterministic and does not require an API key. It shows the agent receiving a task, using a tool, and returning a result.
To try the same agent with a real model:
export OPENAI_MODEL="your-model"
export OPENAI_AUTH_TOKEN="your-token"
# Optional for an OpenAI-compatible endpoint:
export OPENAI_BASE_URL="https://your-provider.example/v1"
uv run python -m scripts.run_bash_agent_demo \
--provider openai \
--task "Inspect this repository and explain what it is for."See .env.example for the supported provider settings.
- Learn how an agent behaves by running and modifying a complete example.
- Build focused agents for coding, research, teaching, or internal workflows.
- Explore tasks that require many steps, tool calls, checks, and revisions.
- Compare prompts, models, tools, and agent strategies under repeatable tasks.
- Evaluate agent behavior on small experiments or established benchmarks.
- Keep the agent easy to understand.
- Prefer useful behavior over impressive abstraction.
- Make long-running work observable and verifiable.
- Make experiments reproducible.
- Add complexity only when it earns its place.
- Runnable demos and experiments
- Evaluation suites and benchmarks
- Project documentation
- Contributing guide
Simple Long Horizon Agent is early-stage and actively evolving. It is designed for learning, research, and small-team experimentation rather than production-scale infrastructure.
Contributions that make the agent simpler, more effective, or easier to learn from are welcome. Read CONTRIBUTING.md and run the local quality gate before opening a pull request:
bash runs/dev/run_ci.shLicensed under the Apache License, Version 2.0.
Footnotes
-
We use a chained workflow similar to ChainSWE, augmented with the
tasktool. ↩ -
We evaluate Qwen3-4B-Base on AIME 2025, BFCL, GSM8K, and HumanEval. Normalized rewards follow the OpenAI PostTrainBench Lite method. ↩