InternScience/InternAgent

InternAgent‑1.5 – A Unified Agentic Framework for Long‑Horizon Autonomous Scientific Discovery

What it is – InternAgent‑1.5 is an open‑source, multi‑modal autonomous‑agent system designed to run end‑to‑end scientific research cycles. It can generate hypotheses, design and execute (simulated) experiments, optimise algorithms, and synthesize literature‑based answers, all without human intervention. The project targets long‑horizon tasks across physics, biology, earth science and life‑science domains.

Key components

Component	What it does
Memory module	Persists experiment outcomes across runs, letting the agent avoid dead‑ends and build on prior successes.
Deep Research (DR) pipeline	Breaks a research question into sub‑tasks, pulls data from academic databases/web, and produces a structured answer or report.
MLEvolve	The optimisation core for algorithm‑design tasks; the authors claim state‑of‑the‑art results on the MLEBench benchmark.
Task suites	tasks/ – algorithm‑discovery problems; sci_tasks/ – paper‑reproduction challenges (e.g., reproducing results from a published astronomy paper).
Launchers	launch.py (unified entry point), launch_discovery.py (full discovery loop), launch_qa.py (question‑answer mode).

How to get started

# 1. Create an isolated environment
conda create -n InternAgent python=3.11
conda activate InternAgent

# 2. Install dependencies
pip install -r requirements.txt

# 3. Provide API keys (OpenAI, OpenRouter, Anthropic, etc.)
cp .env.example .env   # then edit .env with your keys

Running the first demo – the repository ships a toy task called AutoDebug that needs no data download:

python launch_discovery.py \
    --config ./config/default_config.yaml \
    --task AutoDebug \
    --exp_backend claudecode

If you just want a literature‑based answer:

python launch_qa.py --question "What are recent advances in memory‑augmented LLMs?"
# or save to a file
python launch_qa.py -q "..." -o answer.md

Both commands are also reachable via the unified launch.py script.

Configuration – All settings live under config/. default_config.yaml is the starter; you can point the agent to existing idea files (--skip_idea_generation --idea_path …) or swap the model gateway to OpenRouter by editing config/openrouter_config.yaml.

Result handling – Experiments, generated ideas, and logs are automatically written to results/ and logs/ respectively, making it easy to audit the autonomous workflow.

Advanced usage

• Memory persistence – enable via docs/memory_module.md; the agent will read/write a JSON store that survives restarts.
• Deep Research – configure parallel literature crawlers and synthesis strategies in docs/deep_research.md.
• MLEvolve – the optimisation backend for algorithm‑design tasks; see the separate repo InternScience/MLEvolve for details.

Publications & citations – The authors provide two arXiv papers (InternAgent 1.0 and 1.5) and several related works (FlowSearch, AutoMLGen). Use the supplied BibTeX entries when referencing the framework.

Where to find more

• Website: https://discovery.intern-ai.org.cn
• HuggingFace collection: https://huggingface.co/collections/InternScience/internagent
• Technical report: https://huggingface.co/papers/2602.08990

---

Bottom line – InternAgent‑1.5 is a full‑stack autonomous research agent that combines LLM‑driven reasoning, memory‑augmented planning, and optimisation sub‑modules to tackle scientific discovery tasks. The repository includes ready‑to‑run demos, clear configuration files, and documentation for extending the system to new domains or custom experiments.