From the NVIDIA edition of October 22, 2026
deepset and NVIDIA publish a reference architecture for self-improving agents in air-gapped environments, with proposed changes gated by evaluation, approval, audit and rollback
deepset; NVIDIA · AI agent software (governments and regulated organisations) · Europe launch (GTC Berlin); global
- Publisher
- PR Newswire (deepset)
- Original publication
- 2026-10-21
- Source retrieved
- 2026-10-23
- Event date
- 2026-10-22
What happened
deepset, whose Haystack platform underpins the European Commission's AI@EC, released an open reference architecture with NVIDIA for self-improving agents in locally controlled and air-gapped environments. One Haystack agent performs the task; a second analyses traces and proposes changes to the agent harness, each evaluated against predefined objectives and the current version before promotion, with approval, audit and rollback controls. Components are Nemotron models via NIM, NeMo Gym for optimisation, and NemoClaw blueprints running the agent inside OpenShell with policy controls over files, networks, credentials and tools. The first demonstration is a cybersecurity agent analysing confidential source code; a companion GitHub repository and a DGX Spark log-analysis walkthrough are provided.
Why it matters
Relevant to state and university security operations and to any agency considering agents that change their own configuration: it names concrete control points (evaluation gates, approval, rollback) that a U.S. public buyer would need before allowing an agent to self-modify. No public-sector deployment is reported.
Evidence and measured results
Vendor release describing an architecture and demo; no measured improvement, accuracy or false-positive data. The 'self-improving' loop is described, not evaluated. Related OpenShell containment limits were archived Oct 21 (configuration-dependent).
Limitations and uncertainty
Vendor claims; demo only; no independent evaluation; optimisation objectives could reward the wrong behaviour if poorly specified.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-23; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Security teams need help reviewing sensitive code and logs without sending them to external AI services.
- Stakeholders
- State or campus CISO
- Security operations lead
- Change advisory board
- Discovery questions
- Which bounded, read-only task would you pilot?
- Who would approve agent configuration changes?
- Can you operate air-gapped model updates?
- Value hypothesis
- Interpretation: a locally run, governed agent could assist analysts on code or log review; benefit unmeasured.
- Potential engagement
- Governed pilot design for a local security-analysis agent with approval and rollback controls.
- Unsupported claims
- Measured accuracy or analyst time savings
- Safe autonomous self-modification
- Production public-sector use
Pre-sales engineering
Role takeaway
- Fit
- Organisations with confidential code or logs and local GPU capacity.
- Architecture and integration
- Haystack task and optimiser agents, Nemotron via NIM, NeMo Gym evaluation, NemoClaw in OpenShell on local NVIDIA hardware.
- Prerequisites
- Local GPU capacity
- Evaluation dataset of real cases
- Change-management integration
- Constraints
- Demo-stage
- Objectives must be well specified
- NVIDIA runtime stack dependency
- Security
- Enforce least-privilege OpenShell policies and log all agent actions and promoted changes.
- Proof of value
- Compare agent findings with analyst baseline on historical cases
- Test rollback of a promoted change
- Red-team the sandbox policy
Delivery
Role takeaway
- Work
- Deploy locally, define evaluation and approval flow, pilot on a read-only task.
- Dependencies
- Hardware
- Security team time
- Evaluation data
- Ownership
- Security operations, with change advisory board approval.
- Skills and adoption
- Analysts need training to review agent findings and proposed changes.
- Governance checkpoints
- Policy review before pilot
- Approval for each promoted change
- Monthly evaluation review
- Acceptance criteria
- Proposed: precision and recall against analyst-labelled cases
- Proposed: every promoted change has recorded evaluation and approver
- Proposed: rollback executed within agreed time in test
- Risks
- Optimiser reinforcing wrong behaviour
- Over-permissive sandbox
- Reviewer fatigue
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Two-agent loop (task agent plus optimiser agent) with an evaluation environment; runs on local NVIDIA hardware down to DGX Spark; OpenShell for isolation.
Governance
Who approves, reviews and stays accountable for outcomes?
Makes change management of agent behaviour explicit: proposed harness changes need evaluation and human approval before production.
Security and privacy
What data, permissions and controls need testing?
Keeps confidential code and logs local; agent permissions enforced by OpenShell policy, whose effectiveness depends on operator configuration.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Shifts analyst work toward reviewing agent-proposed changes; requires staff able to judge evaluation results.
Procurement
What should contracts, pricing and exit terms secure?
Open-source components reduce licence lock-in for orchestration, but the model-serving and runtime stack is NVIDIA's.
Operating model
Which teams own the service once it runs?
Introduces a release process for agent harness changes similar to software change control.
Publication history
- 2026-10-22NVIDIA · Issue 479 resources
Stable resource ID: deepset-self-improving-sovereign-agents-reference-architecture-nvidia-20261021