# facebookresearch/ReAgent

A platform for Reasoning systems (Reinforcement Learning, Contextual Bandits, etc.)

Repository: https://github.com/facebookresearch/ReAgent
Canonical: https://ross.abutalabs.com/products/facebookresearch-reagent
Homepage: https://facebookresearch.github.io/ReAgent/
Language: Python
License: BSD-3-Clause
License Family: permissive
Last push: 2026-08-24T23:15:31+00:00

## Health v2 (maintenance only)
Score: 77/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 35, longevity 100
- inputs: {"age_days": 3324, "days_push": 9, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 3712, forks 529 (observed 2026-08-28T04:08:15.232068+00:00)

## What it is
ReAgent is an open-source end-to-end platform for applied reinforcement learning developed at Facebook, built in Python with PyTorch for training and TorchScript for serving. It supports deep RL algorithms, contextual bandits, and counterfactual policy evaluation for large-scale offline decision-making tasks. The project is officially archived and no longer maintained, with users directed to its successor Pearl.

## Use cases
- train deep reinforcement learning policies on logged batch data
- run contextual bandit experiments like LinUCB and Thompson sampling
- evaluate new policies offline with counterfactual estimators before deployment
- build RL-based recommender systems with SlateQ or Seq2Slate
- serve trained RL models in production via TorchScript

## When to choose
- you need an offline/batch RL platform for large-scale recommendation or optimization without a simulator
- you want both bandit algorithms and deep RL in one framework
- you need counterfactual policy evaluation (doubly robust, MAGIC) on logged data

## When to avoid
- you need an actively maintained library - ReAgent is archived and unsupported
- you want production-ready RL tooling today - use its successor Pearl instead
- you need online RL with a simulator environment rather than offline batch training

## Facets
- artifact type: framework
- maturity: abandoned
- function: machine-learning, reinforcement-learning, benchmarking
- domain: reinforcement-learning, machine-learning, large-language-models
- platform: python
- tags: pytorch, contextual-bandits, counterfactual-evaluation, recommender-systems, offline-rl, torchscript, facebook-research, archived, linux, docker, gpu

## Member repositories
- facebookresearch/ReAgent (main) score 77

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:08:15.232068+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T18:30:54.231868+00:00, confidence not recorded.
  - readme: https://github.com/facebookresearch/ReAgent (fetched 2026-08-28T04:08:15.232068+00:00, sha 4f33c755d40c)
  - homepage: https://facebookresearch.github.io/ReAgent/ (fetched 2026-08-29T09:24:20.771856+00:00, sha 239134db8fb5)
  - registry_pypi: https://pypi.org/pypi/reagent/json (fetched 2026-08-29T09:24:20.775152+00:00, sha 224cfbfacfa1)
- Data as of 2026-08-30T08:39:29.467469+00:00.