The governed execution runtime for AI agents.
Deterministic by default. AI only when needed.
Nia decides what an AI agent is allowed to do before it acts. Every action is capability-declared, condition-gated, dry-run-validated, and fully audited — so autonomous systems stay reviewable, reproducible, and safe.
Most AI systems can take actions without proving they should.
They wire a model to your tools and hope. There's no record of what was allowed, no gate before the model acts, and no way to replay what happened. That's a problem for a business — and a non-starter for autonomous, multi-agent, and regulated environments.
Nia enforces execution rules before a model is allowed to act.
Actions are declared, not improvised
An agent can only touch the systems its manifest names. Capabilities are written down and reviewable before anything runs.
Models run on a leash
The expensive, unpredictable part — the LLM — is invoked only when an explicit condition holds. Everything else is deterministic.
Every run is evidence
Each invocation leaves a complete audit trail you can read in milliseconds. Reproducible, inspectable, defensible.
The execution model
Four words. One governed pipeline.
Nia's vocabulary is the ideology, encoded into the grammar itself. A worker runs as a sequence of deterministic actions, and reaches for judgment only when it has to.
Worker
A manifest declares the trigger, the permitted capabilities, and the ordered steps. The control plane, in plain YAML.
Run
Each invocation is a Run — isolated, timestamped, and recorded from first action to last.
Action
Deterministic steps execute against your services. No model, no ambiguity, no token cost.
Judgment
Only when a declared condition holds does a Run escalate to a model. Then it returns to deterministic ground.
Deterministic by default. AI only when needed.
CapabilitiesAssurance primitives, not agent features
The controls that make business automation reliable are the same ones autonomous and multi-agent systems require. Nia enforces them at the runtime layer.
Capability declarations
Every worker declares, in its manifest, exactly which systems it may touch. Nothing runs that wasn't written down first.
Condition gating
A model is never invoked unless a declared condition evaluates true against the run's own prior results. Judgment is deliberate, not reflexive.
Dry-run validation
Execute a worker for real with every side effect mocked. If you can dry-run it, you can audit it. If you can audit it, you can trust it.
Per-run audit trail
Every run persists a complete, inspectable record — actions, conditions, judgments, timings, failures. No black boxes.
$ nia inspect worker <name>Read a worker's mind in 200ms — trigger, capabilities, conditions, and the last runs with timings and failures.
$ nia dry-run worker <name>Run it for real with every side effect mocked. Vet a worker before it ever touches a live system.
Where it appliesExercised in the real world. Built for what's next.
The same capability and condition controls that keep commercial automation reliable become essential as AI systems grow autonomous and multi-agent.
Running in production at Caipher across these workloads.
- ✓Customer operations
- ✓Internal business automations
- ✓Lead qualification & follow-up
- ✓Scheduled reporting & ops
Domains where governed execution is a requirement, not a nicety.
- →Research organizations
- →Public sector
- →Critical infrastructure
- →Government programs
Listed for relevance — not a claim of current customers in these sectors.
ProofNot a thesis on a slide. A runtime in use.
Running in production
Caipher operates Nia 24/7. Every release ships after running against real workloads first — dogfooding no competitor can fake.
Open source, AGPL
The deterministic core and the manifest grammar are public. Read the gate, read the audit format, run it yourself.
The agent runs on the runtime
Caipher's own agent executes through Nia — its actions pass the same capability and condition gates as everything else.
github.com/theblockchainbaby/nia
AGPL-3.0 · Runs in production at Caipher, 24/7.
Research & governanceFrom single-machine workers to multi-agent assurance
Nia began as the runtime powering production AI automations at Caipher. The controls that make those automations reliable — capability restrictions, execution conditions, dry-run validation, and full auditability — are the same controls autonomous and multi-agent systems will require.
Our research direction extends Nia's capability and condition model from single-machine declarative workers to multi-agent and multi-tenant deployments — where the assurance guarantees still hold, but production constraints (concurrency, remote execution, multi-principal authorization) are not yet solved.
FAQQuestions, answered straight
Let's talk about governed execution.
For research collaborators, program partners, and teams deploying autonomous AI: reach out for a technical walkthrough of the runtime, the audit model, and where it fits your environment.
Or email support@caipherai.com