Colosseum SWARM Hackathon 2026
The Execution Firewall for
Autonomous AI Agents
Your agent knows what it wants. The chain has a different reality.
Sudont verifies intent against physics — before value moves on-chain.
How It Works
The Diamond Pipeline
Every transaction passes through four nodes. No single node can approve alone. The Judge requires both intent analysis and simulation results to produce any verdict.
Constitution
Policy enforcement
Cortex
Intent extraction
Cage
Physics simulation
Judge
Verdict engine
Core Capabilities
Diamond Architecture
Four-node pipeline: Constitution defines law, Cortex extracts intent, Cage simulates physics, Judge renders verdict. Fan-out / fan-in with cryptographic attestation.
Interrogation Protocol
When a transaction is suspicious but not malicious, Sudont teaches the agent to comply. Structured adversarial feedback loop with immutable audit trail.
Multi-VM Support
First-class EVM (revm) and SVM (LiteSVM) simulation. Single SimulationEngine trait — the Judge is completely VM-agnostic.
TypeScript SDK
Install @sudont/sdk, initialize with vmType, send your first firewall-protected transaction in minutes. Full interrogation handling included.
The Oracle Problem
Approved Destination ≠ Safe Outcome
Between the moment an agent plans a transaction and the moment it lands on-chain, the world moves. Price shifts. Liquidity dries up. Pending state invalidates assumptions. Sudont is the only system that verifies outcomes, not just destinations.
Read the Full Problem Statement →