# Multi-agent enterprise assistants with Voir

> Voir builds multi-agent assistants on governed data. Its Agent Meta-Harness assembles each agent on every turn: it fetches tools from live connectors, wraps them in governance guards, retrieves knowledge and streams the result. Governance is enforced in code inside each tool call, never left to the model.

Source: https://datanature.ai/solutions/multi-agent-enterprise-assistants
Publisher: Dlytica Inc. (DataNature is a Dlytica company)
Updated: 2026-11-05

Governed agents that answer, act and hand back for approval

## Components

- **Voir Agent Meta-Harness**: Builds and runs agents from a declaration of intent and scope
- **Skills**: Reusable, governed playbooks agents call when needed
- **Connectors**: Live links to data and business systems, each isolated with its own timeout
- **Semantic Layer**: Certified metrics agents query instead of raw tables
- **DocuContext knowledge base**: Documents and call audio for retrieval
- **Atlas**: Scopes, budgets, logs and human-in-the-loop approvals

## Steps

1. **Declare the agent**: Define intent, scope, allowed tools, data partitions and SQL allowlists.
2. **Compose per turn**: The meta-harness fetches tools, applies guards, composes the prompt and retrieves knowledge.
3. **Enforce in code**: Document-partition and SQL-allowlist boundaries are checked inside the tool call before anything reaches a downstream system.
4. **Disclose progressively**: Agents start with compact context and pull skills or detail only when needed.
5. **Hand back**: Actions that change systems go to a person for approval when the scope requires it.

## Documentation

- [Agent Meta-Harness](https://docs.datanature.ai/docs/developers/ai-and-analytics/agent-meta-harness)
- [Agents for business users](https://docs.datanature.ai/docs/business/ai-and-analytics/agents)

## How is this different from a chatbot?

A chatbot answers. Voir agents also act: they call systems, follow your procedures and hand work back for approval, all within a defined scope.

## What happens if one system is down?

Each connector is fetched with its own timeout, so one slow or failing system degrades gracefully instead of breaking the agent.

## How do we stop an agent seeing data it should not?

Scope is enforced in code inside every tool call: document partitions and SQL allowlists are checked before the request leaves the platform. The model's cooperation is never assumed.

## Can several agents work together?

Yes. Specialist agents, such as a KYC agent, a credit analyst and a compliance reviewer, can each be scoped and orchestrated within one workflow.

## Common questions

### Why do AI agents fail in production?

Mostly systems problems: excessive permissions, prompt injection, loops and poor data. See [why AI agents fail](/blog/why-ai-agents-fail-in-production).

### How do we stop prompt injection?

Least privilege, approval gates and treating retrieved content as untrusted. See [prompt injection defences](/blog/prompt-injection-defence-for-enterprise-agents).

### Which LLM should agents use?

Cohere, self-hosted open-weight or Dlytica-managed models, chosen per use case. See [three ways to run LLMs](/blog/self-hosted-vs-cohere-vs-dlytica-managed-llms).

## Implementation details

1. **Configure the AI gateway** to your approved model endpoint (self-hosted, Dlytica-managed or an approved provider).
2. **Model the numbers first**: certified cubes over Gold tables, with Cube Security predicates per role.
3. **Load knowledge** into DocuContext partitions that mirror your access model, and check extraction confidence on a sample.
4. **Create one twin per audience** (branch staff, relationship managers, compliance), each with integrations (lakehouse, semantic layer, documents) and a data scope of "Selected only".
5. **Write skills** with worked examples of good questions and answers.
6. **Register external tools** as MCP servers under Connectors, and review the tools each exposes.
7. **Turn on PII guardrails** and require human approval for any action that changes a system.
8. **Test refusal**: ask each twin questions it should not answer, and confirm the boundary holds in the tool call.

The Agent Meta-Harness rebuilds each agent every turn: it fetches tools from live connectors, wraps them in guards, composes the prompt, retrieves knowledge and streams the answer.
