Forge
RL environments for training AI agents on real-world apps
$ forge run gmail --agent anthropic:claude-haiku
▸ Loading Gmail environment (34 emails, 19 contacts)...
▸ Running episode — step 12/50...
▸ Evaluating reward function...
✓ Reward: 0.91 · Policy: passed · Export ready
✓ Dataset saved → sft_pairs_episode_042.jsonlTraining AI agents requires realistic environments
Teams building AI agents struggle to collect quality training data. Synthetic generation misses real app complexity; scraping is inconsistent and brittle. There's no standardized way to create controllable, observable environments where agents can explore, fail, and improve.
Environment Types
Gmail, Slack, CLI, Browser, ready out of the box
Export Formats
SFT, DPO, GRPO, failure datasets, and more
Agent Adapters
Random, scripted, Anthropic, OpenAI, vLLM
Everything you need to collect data
Premade Environments
Gmail replica (34 emails, 19 contacts) and Slack replica (7 channels, 88 thread replies) ready to run, no config needed.
Custom Environments
Describe any app in plain English; 5 parallel LLM agents scaffold code, telemetry, state bridge, policy DSL, and reward function.
Reward Engine
LLM-as-judge, sentence embeddings, ROUGE-L, and BLEU scoring with 6 built-in verifier types (Exact, Event, Temporal, Negative, Policy, Semantic).
Dataset Export
7 formats: SFT pairs, preference pairs, GRPO rollouts, failure dataset, raw trajectories, rewards, verifier results.
Episode Replay & Branching
Re-run any recorded episode step-by-step; fork from any step and try alternate action sequences.
Benchmark Suite
Fixed task suite across 2 domains (email, project management), 10 tasks, scores environments on coverage, reward density, dead-end rate, and action diversity.