Controlled execution for every AI agent NIA ships
The NIA Universal AI Harness is the engineered operating environment around a model. It controls what the model can see, do, remember, call, claim and repeat, and turns an unconstrained model call into an observable, policy-bound run with an independently assessed outcome.
A hybrid harness: a portable embedded runtime per product, a shared control plane publishing signed policies, module manifests, capability packs and evaluator definitions, and shared evidence, audit and telemetry backends. The core is a set of versioned contracts, not a vendor-specific agent framework.
- 9
- Capability level, of 10
- 3
- Planes of control
- 15
- Architecture domains
- 12
- Pluggable interfaces
- 5
- Evidence-based verdicts
- 24
- Control tests
- 4
- Deployment modes
Four design commitments make agent work governable, portable and provable
One foundation, many products
Each product embeds a portable runtime. A single signed control plane governs IntelliStudio and future solutions, so governance is never rebuilt per product.
Authority sits outside the model
Policy is enforced in code at the action boundary, immediately before any side effect. The model reasons; it never authorises.
Evidence before verdict
An independent evaluator issues one of five verdicts against an outcome contract. Missing proof yields INSUFFICIENT, never VERIFIED.
Humans hold material decisions
Approvals are authenticated state transitions that expire. Silence is never consent, and a changed action voids the approval.
IntelliStudio submitted the work. The harness let three tools run, stopped one, and an independent evaluator verified the result.
Reference run for customer ticket T-123. This is the control IntelliStudio is built to enforce. It is not a client's private record.
Ticket T-123 resolved for the correct customer
| Action | Decision | What happened |
|---|---|---|
| knowledge.search | allowed | Ran inside scope |
| crm.customer.read | allowed | Ran inside scope |
| ticket.comment | allowed | Ran inside scope |
| refund.issue | approval | Held until a person approved, then ran |
| credential.export | denied | Stopped before any side effect. The tool did not run. |
Every request passes through three planes before a verdict is issued
Select a component to inspect it, or play a governed run end to end using the blueprint's reference workflow: resolving customer ticket T-123.
Source: AI Harness Build Blueprint v1.0, §3 Reference architecture and §8 Tool call flow; NIA Universal AI Harness Architecture Master Prompt, "The target system in one view".
Each plane answers one question, which keeps authority, execution and proof apart
These are logical responsibilities, not mandatory microservices. A modular monolith is acceptable; separate services must be justified by scale, isolation or deployment constraints.
Control plane
Holds the rules, the approved agents and the approved versions. It publishes signed bundles; runs continue on the last valid bundle if it goes offline.
- Agent registry, identity and versioning
- Lifecycle, tenant boundaries, RBAC/ABAC
- Policy and permissions, approvals, budgets
- Signed module, policy and capability registry
Execution plane
Runs the agent loop: builds context, calls the model, routes every tool call through a gateway and stops for a human at material boundaries.
- Agent loops and orchestration
- Context builder, memory, tool gateway
- Model adapters, retries, sandbox, queues
- Human gates as state transitions
Evidence and assurance plane
Records tamper-evident evidence and has an independent evaluator decide whether the outcome was achieved. A self-report never counts as proof.
- Audit events, traces, metrics, evals
- Artifact provenance and outcome receipts
- Drift, regressions, replay and simulation
- AgenticAssure attestation interface
Fifteen domains set the standard every component is measured against
Ten agent-engineering disciplines and five enterprise requirements. Open a domain to see its practices and locate it on the system map.
Source: NIA Universal AI Harness Architecture Master Prompt, "The 15-domain architectural standard".
Every run follows a legal path, and only sufficient evidence can complete it
Execution and evaluation are separate states. Human approval is a state transition, not a chat prompt, and missing evidence can never be promoted to success.
| Verdict | Meaning | Completes |
|---|---|---|
| VERIFIED | All mandatory criteria have sufficient permitted evidence | Yes |
| PARTIAL | Mandatory criteria pass; optional criteria unresolved | By policy |
| INSUFFICIENT | Required proof unavailable, stale or inadequate | No |
| CONFLICTING | Qualifying evidence supports incompatible states | No |
| FAILED | A mandatory criterion is proved false | No |
Precedence: FAILED › CONFLICTING › INSUFFICIENT › VERIFIED. Confidence describes uncertainty; it cannot promote INSUFFICIENT to VERIFIED.
| ID | Scenario | Required behaviour |
|---|
Source: AI Harness Build Blueprint v1.0, §6 Event model and §9 Evaluator loops and verdicts; Master Prompt, "Required test matrix".
Phases 0 to 4 are delivered, and Bands 5 to 9 take the harness to level 9
On a scale where blockchain is 1 and AGI is 10, Phases 0 to 4 land the harness at about 5. Bands 5 to 9 are the climb, and the harness stops short of AGI: every run has an outcome contract a human published, and there are no unattended production weight writes. Status is reported from repository evidence only.
| Mode | Use | Control-plane dependency |
|---|---|---|
| Embedded library | Local development, CLI, low-latency service | Optional |
| Sidecar / runtime service | Several applications share a host | Periodic |
| Kubernetes worker | Scalable asynchronous workloads | Periodic or event-driven |
| Air-gapped bundle | Regulated or restricted environment | Offline signed export |
Source: Master Prompt, "Phased execution instructions"; AI Harness Build Blueprint v1.0, §3 Deployment modes.