Quick Start
Spin up the environment and run your first agent loop in minutes.
Action Space
Explore all 8 action types your agent can execute.
API Reference
Full REST endpoint reference with request/response schemas.
Task Catalogue
Three difficulty tiers — Easy, Medium, and Hard — with escalating complexity.
How It Works
The environment runs as a FastAPI server. Your agent interacts with it through three core endpoints —/reset, /step, and /state — and submits its cleaned dataset when ready. A hidden grader engine scores the final output against ground truth.
1
Start the server
Run the environment locally or via Docker. It listens on port
8080.2
Reset to begin an episode
POST /reset initializes a fresh episode, loads messy CRM data, and returns your first observation with the task objective and available data sources.3
Execute cleaning actions
Your agent sends
POST /step with a CRMPipelineAction payload — view data, profile columns, standardize formats, handle missing values, deduplicate records, or run SQL joins.4
Submit and get scored
Send
SUBMIT_PIPELINE to trigger the grader. It compares your final dataset against hidden ground truth and returns a score from 0.0 to 1.0.Task Difficulty Levels
Task 1 — Easy
Normalize a single
web_forms dataset. Fix dates, lowercase emails, strip spaces.Task 2 — Medium
Merge
web_forms and legacy_db. Handle schema mismatches and deduplicate records.Task 3 — Hard
3-way merge across Salesforce, Web Leads, and Legacy DB with conflict resolution rules.
Key Capabilities
Architecture & POMDP Model
Understand the environment’s state machine, reward signals, and data flow.
Writing an Agent
Step-by-step guide to building a custom agent that interacts with the pipeline.
Data Engine
Deep-dive into data sources, generation logic, and all standardization strategies.
Grading & Scoring
Understand exactly how your agent’s output is evaluated and scored.