AVM

AVM is a runtime that puts agent tool calls inside a verifiable execution environment: models submit tool calls over MCP, the runner executes in an isolated sandbox or container and returns signed output, and the caller can check that execution was not tampered with. Resource governance uses a gas budget model that halts a call on overrun so a single request cannot drain a quota. Sandboxes are destroyed after a call and environments are never reused across requests.

Security runs two filters: injection scanning matches prompt attack patterns against 91 rules, and privacy classification uses 27 patterns plus a Luhn check to catch card numbers and similar data. Execution is described by 25 opcodes and can produce a zero-knowledge proof on demand, making the claim that this output came from that program on that input verifiable without disclosing the input. Opcodes cover reads, writes, branching, external calls and proof generation, and proofs can be verified by a contract onchain or checked directly by a counterparty offchain.