CAAF: A New Framework for Building Deterministic AI Agents in Safety-Critical Domains

This article introduces the Convergent AI Agent Framework (CAAF), which shifts AI agents from open-ended generation to closed-loop safety and determinism through recursive atomic decomposition, a unified assertion interface, and state locking mechanisms. It achieves 100% paradox detection in autonomous driving and pharmaceutical domains, providing a reliable solution for safety-critical systems.

Large language models (LLMs) perform well in general tasks, but there is a fundamental controllability gap in safety-critical domains: even a low rate of undetected constraint violations makes deployment impossible. Core issues include sycophantic compliance (catering to users instead of strictly enforcing safety constraints), context attention decay, and random oscillations—these can lead to catastrophic consequences in scenarios like autonomous driving and pharmaceuticals.

Pillar 1: Recursive Atomic Decomposition and Physical Context Firewall

Split complex tasks into indivisible atomic operations, clarify physical context boundaries, ensure sub-task specifications are clear, constraints are explicitly encoded, and irrelevant information is isolated.

Pillar 2: Unified Assertion Interface (UAI)

A core innovation that formalizes domain invariants into a machine-readable registry, enabling deterministic execution and real-time interception of violations instead of post-hoc verification.

Pillar 3: Structured Semantic Gradient and State Locking

Ensure monotonic convergence through state locking to prevent the system from reverting from a safe state to an unsafe one; semantic gradients provide fine-grained control over state transitions.

Autonomous Driving (SAE Level 3)

Under 30 test cases and 7 conditions, CAAF-all-GPT-4o-mini achieved a 100% paradox detection rate, while the standalone GPT-4o (temperature 0) had a 0% detection rate.

Pharmaceutical Continuous Flow Reactor Design

In scenarios with 7 constraints and nonlinear Arrhenius interactions, CAAF still maintained a 100% detection rate, and the Mono+UAI ablation experiment reached 95%.

Multi-agent Comparison

Architectures like debate and sequential checking had a 0% detection rate, confirming that UAI is the core of reliability.

CAAF successfully reveals that in safety-critical domains, reliability is more important than capability. Its advantages include prompt independence, single-model offline deployment, and formal guarantees via UAI. CAAF represents an important shift of AI agents from generative capability to verifiable deterministic behavior.

Autonomous Driving

Provides a safety architecture for L3/L4-level decision systems, alleviating the black-box problem.

Industrial Control

Process industries like pharmaceuticals and chemicals can convert operational specifications into executable machine code.

AI Research

Need to balance capability with controllability, verifiability, and determinism.

Current limitations: Only targets structured constraint scenarios, requiring domain experts to build invariant registries. Future directions: Explore automated constraint extraction, soft constraint mechanisms, and integration with other formal verification methods.