How future-house/robin works
Project type and purpose
• This is an open-source Python framework and CLI (via Jupyter notebooks) called “Robin,” designed to orchestrate a multi‐agent workflow for scientific discovery.
• It provides reusable modules to generate experimental assay hypotheses, literature reviews, therapeutic candidate proposals, and optional data analysis, then writes structured outputs to disk.
File structure
Top-level layout:
• pyproject.toml / uv.lock – dependency and lock files
• robin/ – core Python package
– configuration.py: load API keys, set up RobinConfiguration
– prompts.py: text templates for queries
– assays.py: experimental assay generation
– candidates.py: therapeutic candidate generation
– analyses.py: optional data analysis (Finch agent)
– multitrajectory_runner.py: high-level orchestrator for multi‐step flows
– utils.py: common helpers, I/O routines
• robin_demo.ipynb & robin_full.ipynb – user-facing notebooks to run workflows
• examples/ – static sample outputs for 10 diseases, demonstrating folder structure and file types
• robin_output/ – the runtime output directory created per run (timestamped disease subfolder)
• .github/ & CI configs – build and test automation
README insights
• System relies on environment variables or explicit config for LLM keys (OpenAI, other LiteLLM backends) and FutureHouse API.
• Core workflow:
– Instantiate RobinConfiguration (disease, API keys, model)
– Generate and rank experimental assays (assays.py)
– Generate and rank therapeutic candidates (candidates.py)
– Optionally run data analysis via FutureHouse Finch agent (analyses.py)
– Save outputs: text files, CSV rankings, summary files, nested in robin_output/<disease_timestamp>
a. Main components
1. User Interface
– Jupyter Notebook front-end (CLI wrapper)
– Python REPL / script entrypoint
2. Application Core
– robin.configuration → loads keys, parameters
– robin.multitrajectory_runner → orchestrates the sequence
– robin.assays → Experimental Assay Agent
– robin.candidates → Therapeutic Candidate Agent
– robin.analyses → Data Analysis Agent (Finch)
– robin.prompts → Query templates
– robin.utils → File I/O, logging
3. External Services
– LLM Providers (OpenAI, other via LiteLLM)
– FutureHouse platform (Finch data analysis agent)
4. File System
– examples/ (static examples)
– robin_output/ (dynamic, timestamped run directories)
b. Relationships and interactions
– Notebook → calls Application Core
– Configuration → passed into each agent
– Agents → make HTTP/API calls to LLM providers or FutureHouse
– Agents → receive responses → robin.utils formats responses → writes to robin_output
– Multitrajectory_runner → loops over assays/candidates → aggregates rankings → triggers optional analysis
c. Architectural patterns
– Orchestration pattern: multitrajectory_runner as a coordinator
– Agent pattern: separate modules for distinct responsibilities (hypotheses, literature review, ranking)
– Template method: prompts.py provides reusable text templates
– Plug-in/extensibility: configurable LLM backends via LiteLLM, optional analysis stage
d. Key technologies and libraries
– Python 3.12+
– LiteLLM (abstracts multiple LLM providers)
– requests or httpx for API communication
– pandas (for CSV reading/writing, ranking)
– Jupyter notebooks (user interface)
– CI: GitHub Actions