GenAI & agents

Verifiable Agent Kernel (VAK)

v1.0 · 80%+ test coverage

Deterministic control plane for LLM agents. Least-privilege ABAC, WASM-sandboxed tools, Z3 formal verification before execution, 3-tier memory (Merkle DAG), multi-agent debate with quadratic-voting consensus, and a PRM fine-tuning toolkit. v1.0 with 80%+ test coverage and LangChain/AutoGPT/MCP adapters.

Code

Problem

LLM agents act through tools with real side effects, yet their behaviour is non-deterministic and hard to audit. Teams need hard guarantees about what an agent may do before it does it.

Approach

Least-privilege access control with Cedar/ABAC policies
Tools run inside a WASM sandbox
Z3 formal verification checks each action before it executes
Three-tier memory backed by a Merkle DAG
Multi-agent debate with quadratic-voting consensus, plus a PRM fine-tuning toolkit

Verifiable Agent Kernel (VAK) workflow: 5 stages from input to result — Workflow

Results

v1.0 release with 80%+ test coverage
CI/CD on Docker and Kubernetes
Adapters for LangChain, AutoGPT and MCP

Stack

Rust
Python
WASM
Cedar / ABAC
Z3
Docker
Kubernetes
MCP

By Muzammil Nawaz Khan, AI Engineer.