OpenProgram Docs
Research Agent#
An autonomous research agent: take a research topic and walk the full pipeline of literature survey → idea generation → experiments → writing → review → rebuttal / presentation, producing a submission-ready paper. It does not trust its own output — every citation is verified against four indexes (Crossref / OpenAlex / Semantic Scholar / arXiv), every number in the paper must trace back to an experiment's run_record.json, and the review can run on a different model (author and reviewer use different models to avoid self-grading).
Install#
openprogram programs install research
No extra dependencies — the harness uses only OpenProgram itself. Optional: install pymupdf for PDF parsing (needed by load_paper / wiki ingest), and install the wiki harness for knowledge-base integration (openprogram programs install wiki).
Usage#
The entry function is research_agent, registered as a tool (as_tool=True, toolset research). In chat, just describe a research task to trigger it, e.g. "Survey recent work on LLM uncertainty".
Run it directly from the command line:
openprogram programs run research_agent -a task="Survey recent work on LLM uncertainty"
Internally it is a two-level controller. At the first level, the LLM chooses which research stage to enter (literature / idea / experiment / writing / review / rebuttal / presentation / theory / knowledge / project; stages have dependency ordering, and missing prerequisites are filled in first). At the second level, within a stage, the LLM picks and runs that stage's functions one by one. About 89 functions across 10 stages; each function is an ordinary Python file whose docstring is the prompt, directly editable.
Hidden entry parameters (available to code / CLI callers):
| Parameter | Description |
|---|---|
review_runtime |
When provided, review functions run on a different model (cross-model review) |
work_dir |
Project working directory |
max_runtime_s |
Soft time budget: after the deadline no new work starts; running steps finish and wrap up normally |
stop_event |
Graceful stop signal (any object with is_set()); wraps up after the current step finishes |
The return value is a dict: task, success, summary, stages_completed, history.
Dependency notes#
- Citation verification, uncited-claim checks, citation-dump checks, and similar verifications are pure Python / regex — no extra tokens;
integrity_gateuses a single bounded LLM call. - Online retrieval (arXiv / Semantic Scholar, etc.) requires network access; LaTeX compilation requires a local TeX distribution.
Source and README: openprogram/functions/agentics/Research-Agent-Harness/, upstream repository Fzkuji/Research-Agent-Harness.