The natural
environment
for AI agents
is onchain.
training grounds.
Agentic reinforcement learning environments where AI agents learn to operate onchain through trial and error. No wrappers. No shortcuts. The actual production skill.
AI can spot exploits across 4.6 million historical
attacks.
It writes Solidity fluently.
It can explain concentrated liquidity in textbook prose.
And then —
it cannot swap a token.
We call this the Knowledge–Operational gap. The difference between understanding blockchain concepts and being able to execute blockchain operations. Between knowing what a flash loan is and actually constructing the callback, executing the arbitrage, and repaying it in a single transaction.
AI agents know crypto.
They can't do crypto.
Frontier models write Solidity but cannot execute a DeFi swap, navigate MEV, or react to onchain state. Knowledge without operational capability is worthless onchain.
Evals.
Detection.
EVMbench · SCONE-bench · Codex 5.3 hits 70%+ on bug detection. The category is mature, the question is answered, the room is full.
Live envs.
Execution.
Agents dropped into simulated worlds. Trial-and-error in live forks. Rewards from onchain state. Trains operational capability.
40+ companies build Operational RL environments. Every single one targets coding, computer use, or enterprise workflows.
0 / 40+ → crypto
Five categories.
Zero wrappers.
Each environment ships as a Docker image. Each task ships with a machine-verifiable grader. Agents interact via raw JSON-RPC. No SDK lock-in. No language constraints.
Bring your agent.
We bring the world.
The category is proven.
The crypto vertical is untouched.
Estimated annual spend on RL environments across all frontier labs. The category is real, venture-validated, and growing fast — but the crypto vertical is empty.
Applied Compute went $0 → $1.3B valuation in 8 months. Total RL environment market estimated at $4–8B/year across all labs.
Anthropic and OpenAI each spend ~$1B/year on RL environments. OpenAI projects $8B by 2030. Anthropic uses 12+ RL vendors at $300–500K/quarter.
EVMbench (OpenAI/Paradigm), SCONE-bench (Anthropic). Labs treat crypto as legitimate. But Knowledge RL ≠ Operational RL.
Coinbase Ventures names onchain agent training a core focus. Top crypto VCs identify RL fine-tuning as missing infra.
40+ companies build Operational RL. Zero target crypto. Wrapper-dependent agents (Wayfinder, Clawi) hit ceilings on complex DeFi.
Build
with us.
Pre-seed · 2026. Looking for AI labs to co-develop environments tailored to their training pipelines. Limited slots.