Agentic Research connects evidence discovery, falsifiable hypotheses, experiment planning, real Python or Notebook execution, held-out validation, and research writing in a resumable autonomous research workflow. It treats generated claims as hypotheses to test, not as scientific results.
Provide a topic such as “reliable few-shot medical image segmentation.” Agentic Research collects and organizes evidence, proposes falsifiable hypotheses, and creates an experiment plan with a baseline, candidate approach, metrics, seed schedule, and pass criteria. The plan pauses for your approval before generated code runs.
After approval, Agentic Research generates Python experiments or parameterized Jupyter notebooks and runs them locally under controlled limits. Each trial retains its code, inputs, raw results, logs, exit status, and duration. Executed notebooks are retained for review and reproduction.
Development seeds are used to improve a candidate; separate held-out seeds are used only for the final decision. Agentic Research compares baseline and candidate success rate, minimum improvement, and variability, then reports accepted, rejected, inconclusive, or invalid rather than treating a single successful run as a finding.
Completed runs can produce Markdown, LaTeX, and optional PDF reports. Citations are drawn from the saved evidence snapshot, and results and limitations are included. Unvalidated results are explicitly labelled instead of being presented as positive conclusions.
Use the Tauri desktop app, CLI, or Web console to see progress, live logs, metrics, hypotheses, reports, and approval requests. Interrupted runs can be resumed or cancelled. The desktop UI supports Simplified Chinese and English, follows the OS language on first launch, and remembers the user's explicit choice. Its appearance can follow the system or be set to light or dark. Desktop data stays in the OS application-data directory; Web/CLI artifacts remain under data/workspace/.
Model execution is unified on the OpenAI Agents SDK. OpenAI uses the native Responses or Chat Completions model path; Anthropic Claude and Google Gemini use the SDK's LiteLLM adapter. Each provider still has its own Base URL, model ID, and API key.
- Metrics, thresholds, and seed schedules are fixed before each experiment; baseline and candidate are compared under the same conditions.
- Generated code does not run through a shell. API keys are removed from its environment, and execution has time, memory, and log limits.
- Agentic Research helps design, execute, and audit research; it cannot prove that a study design is sound, data is leak-free, or a conclusion is statistically significant. Domain review remains necessary for high-stakes research.
Research direction
-> multi-source evidence retrieval and deduplication
-> falsifiable hypothesis and structured review
-> baseline/candidate experiment plan
-> human approval by default
-> bounded iteration on development seeds
-> independent held-out validation
-> accepted / rejected / inconclusive / invalid
-> Markdown / LaTeX / optional PDF report
Python 3.10+ and Node.js 20+ are required. Install dependencies only through the official package-manager commands:
uv sync --extra dev
npm --prefix frontend ci
cp .env.example .envDesktop development additionally requires Rust stable and the Tauri 2 system prerequisites:
npm install
npm --prefix frontend installThe desktop settings page can save provider credentials to its local, Git-ignored .env. In cloud mode the page is read-only: a server administrator supplies provider-prefixed environment variables so regular web users cannot modify global credentials. Keys are never returned by the API.
The defaults are gpt-5.6-terra through the Responses API, claude-sonnet-5, and gemini-3.5-flash. For a third-party compatibility gateway, use only model IDs and API modes exposed by that gateway.
Run diagnostics and the real-process offline demo before configuring an external model:
uv run python -m backend.cli doctor
uv run python -m backend.cli demo
uv run pytestThe demo launches 12 isolated child processes and stores its auditable artifacts under data/workspace/runs/.
uv run python -m backend.cli run --direction "Reliable few-shot medical image segmentation" --max-ideas 2
uv run python -m backend.cli status
uv run python -m backend.cli approve <run_id>
uv run python -m backend.cli resume <run_id>
uv run python -m backend.cli cancel <run_id>
uv run python -m backend.cli daemonWith the default configuration, a new run pauses at waiting_review before generated experiment code executes.
Build the frontend and start the local API/static server:
./start.shOpen the port configured by BACKEND_PORT (the example .env uses http://127.0.0.1:4019).
The desktop app reuses the existing React/Vite frontend and bundles FastAPI as a Python sidecar, so end users do not need to install Python or Node.js.
npm run desktop:dev
npm run desktop:buildThe build scripts generate the target-triple sidecar automatically. The Rust host chooses an unused loopback port, creates an ephemeral API token, starts the backend, and terminates it when the app exits. Desktop configuration and research artifacts live in the OS application-data directory. See the Chinese desktop architecture guide for details.
cp .env.example .env
docker compose up --buildThe service binds to 127.0.0.1:8000, runs as a non-root container user, and persists data/workspace/ on the host.
- Chinese architecture guide
- Research protocol
- Security boundaries
- Frontend guide
- Tauri desktop architecture
Static checks and local resource monitoring are defense-in-depth controls, not a strong sandbox. Run untrusted generated experiments inside a dedicated container or VM, and require domain review before treating output as scientific evidence.