Agent studio
A visual builder for the whole agent: persona, knowledge, tools, model policy and guardrails, with live preview as you edit and a hard line between what you are drafting and what customers are talking to.
- Draft and live are separate, so an edit never reaches production by accident
- Publish creates an immutable version you can roll back to
- Start from a role template, or from an empty canvas
- Pro-code extension points where the visual builder runs out
Model gateway & catalog
One governed abstraction in front of every provider. The catalog records which deployments are approved, with capabilities, regions, pricing and eval history; the gateway enforces the policy at call time.
- Routing, fallback and retry across providers
- Prompt caching and streaming built in
- Disallowed model usage is blocked, not just discouraged
- Every call emits usage, latency and cost metadata
Knowledge & memory governance
Ground answers in your own material, and govern what each agent is allowed to retrieve. Upload documents, sync a drive, or scrape URLs. Chunking, embedding and indexing happen for you.
- Agents retrieve only the sources they are authorised for
- Knowledge is versioned, with sync and index status visible
- Retrieval quality, freshness and citation coverage are measured
- Knowledge gaps surface as evidence, not guesswork
Enterprise agent registry
A single inventory of every agent in the organisation, the ones built here and the ones built elsewhere. Ownership, dependencies, risk rating and lifecycle stage in one place.
- Register external agents by manifest with scoped credentials
- External agents send health and trace telemetry like native ones
- Filter by owner, risk, stage or dependency
- Programmatic registry management through the API
Projects & environments
Agents are configured per project and per environment. Development, staging and production come with every project, and you can add your own up to five.
- Config, credentials and model policy resolve per environment
- Release candidates pin version, manifest hash and dependencies
- Promotion needs gate evidence, or an audited override
- Deployment orchestration for release, rollback, suspend and retire
Tool execution & connector registry
Agents that do things, not just say things. Every tool, connector, MCP server, API and webhook is catalogued, scoped and approved before an agent can call it.
- Argument validation and rate limits on every call
- High-risk actions pause for human approval
- Tool invocations are traced and version-linked
- Admins control which agents see which tools
ZNYX runtime guardrails
Guardrails run on the ZNYX runtime as the system of record, on the hot path rather than as an afterthought in the prompt. Output is buffered and evaluated before any of it reaches a customer.
- Input, output and tool calls scanned against your policy
- Allow, redact or block, decided server-side
- Fail-closed in production if the policy service cannot answer
- Every guardrail decision lands in the trace
Traceability & observability
Every production turn produces a structured trace: what was retrieved, which prompt version ran, which model answered, what it cost, which tools fired and what the guardrails decided.
- Inspect any turn by trace ID with policy-aware redaction
- Latency, token and error metrics per agent and environment
- Alerts on quality and health regressions
- External agents emit into the same pipeline
Evaluation & promotion gates
Versioned eval datasets, offline runs and human review queues, then a gate that decides whether a candidate is allowed anywhere near production.
- Build datasets from real production samples
- Gates check pass rate, hallucination, safety, citations, latency, cost and regression
- Blocked promotions state exactly which gate failed
- Overrides need an authorised approver and are audited
FinOps for AI
An append-only cost ledger built from real model and trace evidence, so AI spend is attributable to an agent, a project and an environment instead of arriving as one monthly invoice.
- Cost per agent, model, project and environment
- Cost per resolution tied to outcomes
- Budgets with anomaly signals
- Exports for finance
Human-in-the-loop operations
The queue where people and agents meet: approvals for risky actions, escalations when confidence drops, and handoffs that arrive with the whole conversation attached.
- Approval queues for high-risk tool actions
- Escalation on low confidence, guardrail block or policy rule
- Full context handed to the human taking over
- Separation of duties enforced on approvals
Advanced runtime orchestration
When one agent is not enough: multi-agent workflows expressed as a validated DAG, published immutably with a hash chain so you always know which topology ran.
- DAG validation before publish
- Immutable hash-chained workflow versions
- Agents calling agents, still inside the guardrails
- Multi-step task execution across connected tools
AI improvement loop
Production evidence becomes clustered insight, then a reviewable proposal. Gate-first by design: the loop never mutates a production agent on its own.
- Signals clustered into insights, not raw noise
- Proposals reviewed and approved by a person
- Improvements re-enter the lifecycle as a new candidate
- Closed-loop measurement against the original signal
Internal agent marketplace
A place for teams to share what already works: vetted agents, prompt templates and tool configurations, without anyone installing straight into production.
- Publish vetted agents and prompts internally
- Installs project into a draft, never a live agent
- Reviews and production requests are governed
- Templates by role and industry
Admin & developer portal
One surface for administrators to run the platform, and one for developers to build on it: API keys, SDKs, onboarding and an append-only admin audit.
- Scoped API keys per org, project and environment
- Python SDK, with more clients following
- Self-service onboarding for external agent teams
- Federated search across platform resources