White paperGoverning AI agents in production: policies, human approvals and an emergency stop.Read the white paper
NavinAI Platform
FRFrançais
Documentation
Request a demo

AI agents

AI agents that act, under control

The Navin runtime executes long-running, tool-using, multi-step agents. They only reach models, tools, MCP servers and data through the platform gateways.
Request a demoContact us

At a glance

  • Durable runtime with replay
  • MCP isolated per user and per project
  • Human in the loop and kill switch

The challenge

From prototype to production agent

A demo agent fits in a few lines. A production agent must survive failures, respect the user's rights, ask for validation before acting and account for every step.

  • Tools called with too many rights
  • Uncontrolled MCP servers
  • No recovery after an incident
  • No visibility into intermediate steps

The execution model

The agent in the center, the gateways around it

The agent holds no key and no direct access. Every capability goes through its gateway, which enforces policy, DLP and audit.

  • Model Gateway

    LiteLLM, vLLM, remote providers

  • Tool and MCP Gateway

    Tools, internal MCP servers, plugins

Agent runtime

AgentLoop, AgentRunner, RunPlan, sub-agents

  • Knowledge Gateway

    Documents, RAG, context, memory

  • Data Gateway

    Classification, DLP, tokenization

Bypassing an SDK grants no extra access: security is enforced by the gateways, the network and identities.

The runtime

Everything reliable agents need

  • Navin agent runtime

    Agent loop, plan controller, sub-agents, sessions and message bus; optional LangGraph and Temporal adapters.

  • Durable execution and replay

    PostgreSQL checkpoints, recovery after an incident and replay of a run for analysis.

  • Human in the loop

    The agent asks for validation before a sensitive action; the decision and its author are traced.

  • Sandbox

    Code and tools run in a gVisor or Kata sandbox, with resource limits.

  • Isolated MCP

    Every MCP runtime is isolated per user and per project on the ToolHive Kubernetes operator. Only the internal servers of the catalog are allowed.

  • Skills and plugins

    A library per project gathers skills, plugins, MCP servers, channels and rules, versioned and approved.

  • Agent to agent

    Agents collaborate over the A2A protocol, each with its own identity and rights.

  • SDKs and contracts

    Python, TypeScript, Java and Go SDKs; manifests and typed calls validated locally against the platform contracts.

The proof

Agents observable end to end

Every step of a run, every tool call and every policy decision is visible in the portal and in the audit trail.

  • History of runs and their steps
  • Cost and tokens of every run
  • Evaluation of answer quality
  • Immediate stop with the kill switch

Go further

  • DatasheetAgents

    Isolated MCP runtimes per user and project

    The MCP registry and gateway decide who may use which server; the MCP runtime creates isolated instances on demand on the ToolHive Kubernetes Operator and destroys them after use. Every tool call still goes through policy, DLP and approval.

    2 min read
  • White paperGovernance

    Governing AI agents in production

    How Navin Enterprise AI keeps autonomous agents under control: OPA policies stacked from platform to tenant to project, approvals with a human in the loop, a kill switch available at any time and a governance registry as the single source of truth.

    25 min read

The other capabilities

Put your first agent into production

We pick a use case with you and the shortest path to a governed go-live.

Request a demoContact us