deepset unveils reference architecture for self-improving agents
deepset unveiled an open reference architecture on October 21, 2026, for self-improving AI agents designed to operate in locally controlled and air-gapped environments, combining the company's Haystack orchestration framework with NVIDIA open models and software.
Architecture and Technical Stack
The architecture integrates Haystack's modular agent orchestration with NVIDIA Nemotron open models served through NVIDIA NIM microservices, along with NVIDIA NeMo Gym as an optimization environment and NVIDIA NemoClaw open blueprints that run agents inside NVIDIA OpenShell. OpenShell provides policy-based controls over files, networks, credentials, and tools—enabling operators to govern agent behavior at runtime without requiring model retraining.
The design employs a dual-agent feedback loop: one agent performs the domain task while a second agent analyzes execution traces and results, proposing refinements to the agent harness. Proposed changes are evaluated against predefined objectives and the current version before promotion, with built-in approval, audit, and rollback controls. The entire system can run on locally controlled NVIDIA-accelerated infrastructure, including NVIDIA DGX Spark.
First Use Case: Cybersecurity
deepset identified cybersecurity as the initial proving ground for the architecture. The implementation involves an agent that analyzes confidential source code for vulnerabilities and uses feedback from prior runs to refine performance—demonstrating how self-improving agents can operate on sensitive data without transferring it to external systems or cloud infrastructure.
Why This Matters for the Agent Economy
The release addresses a critical gap in enterprise adoption: the need for agents that improve over time while remaining under organizational control. By combining orchestration, model serving, optimization, and governance in a single reference architecture, deepset and NVIDIA are lowering barriers for organizations that require air-gapped deployment—particularly in government, defense, and classified information handling.
The modular design separates the agent logic layer (Haystack) from the infrastructure layer (NVIDIA), allowing enterprises to adopt the pattern without vendor lock-in. The dual-agent feedback mechanism enables continuous performance improvement through evaluation rather than retraining, reducing operational costs and data movement.
This is a developer-facing tool release—the architecture is open and documented for teams building governed agents on-premise. The cybersecurity proof-of-concept validates the approach in a domain where data sensitivity and regulatory constraints make cloud-based agent deployment impractical.
