Home / Products / AgentManager

Arvenza Preview AgentManager Govern every model call your agents make.

AgentManager puts one gateway between your agents and the AI models they use. Every request is authenticated, checked against your guardrails and budgets, routed to the right model and recorded, so you decide who may use which models, at what cost and within which rules.

Arrange a Conversation What's Available Today

Agents keep the chat API and the client libraries they already use. Only the address they call changes.

Requests from three agents pass identity, guardrail and budget checks in the gateway. One goes straight to a model, one fails over when a provider is down, and one is stopped by a guardrail. Supportagent Financeagent Operationsagent AgentManager gateway Identity Guardrails Budget Provider Acloud model Provider Bcloud model Self-hostedyour model
Each request passes identity, guardrail and budget checks. One goes straight through, one fails over when a provider is down, and one is stopped by a guardrail.

A gateway on the request path, a control plane behind it

Operators make decisions in one place; the gateway enforces them on every call and reports back what happened.

The gateway

Sits between agents and model providers. It authenticates each caller, enforces model access, limits, budgets and guardrails, routes with retries and failover, and records a usage event for every attempt.

The control plane

Organisations, projects and environments; agents that are active, pending, suspended or revoked; keys that are issued, rotated and revoked. Publishing turns decisions into gateway configuration, as numbered revisions you can compare and roll back.

The operator console

Spend, traffic and latency at a glance, a trace for every request, usage and cost by agent and model, and pages for guardrails, limits, budgets, credentials, safeguards and audit.

Available today, and what comes next

AgentManager is in preview. Everything in the first list is built and verified by automated tests or live runs. The second list is specified and on the way.

Available today

  • One standard chat-completions API in front of your model providers, streaming included, plus model listing, embeddings, re-ranking, image generation and moderation
  • Accepts both of the most widely used chat API formats
  • Routing by failover, round robin, consistent hashing, lowest cost, lowest latency or least busy, with retries and cool-downs
  • Rate limits across seven dimensions, budgets, and an exact-match response cache
  • Guardrails for keywords, personal data, prompt injection, access control, custom rules and model-based review
  • Policy scripts in Starlang, sandboxed with no access to files or the network
  • Safeguards written in plain language, turned into rules, tested against your own examples, then moved from draft to monitoring to enforcing
  • A usage event for every attempt, metrics, and OpenTelemetry tracing
  • Hot reload that rejects an invalid configuration whole, graceful draining, and offline validation
  • Role-based access with 33 permissions and six roles, single sign-on, request traces and an audit log

On the roadmap

  • Gateways for MCP tool traffic and agent-to-agent (A2A) calls
  • An admin API and distributed configuration
  • Audio, file, batch and realtime endpoints
  • Semantic caching and screening of streamed output
  • Built-in TLS termination
  • Budget alerts, environment promotion and custom roles
  • Container images and orchestrated deployment
  • Automatic connection of agent workloads to the gateway

Why put a gateway in front of your models

Control
Who may use which model
Keys are tied to agents, projects and environments, and each key can reach only the models it is allowed to.
Rules that hold
Guardrails and policies run in the gateway, so every agent that calls through it gets the same rules, whoever built it.

Cost
Budgets and limits
Cap spend with budgets and hold request rates within limits, before a runaway loop becomes an invoice.
The cheaper route
Route by cost or latency, and serve repeated requests from the cache.

Resilience
Failover
When a provider slows down or fails, requests move to the next one after retries, and come back when it recovers.
Changes without downtime
Configuration reloads live; an invalid change is rejected before it can take effect.

Evidence
Every call accounted for
Each attempt produces a usage event, so the gateway's records are the basis for chargeback and review.
An audit trail
Changes to keys, policies and routes go to an audit log, and every request has a trace you can open.

Works with the rest of Arvenza

HelixSet the gateway as the endpoint in Helix's model settings and every agent's model calls are governed. VGLRoute the model calls behind VGL Studio's query drafting through the same gateway. Arvenza DBKeep retrieval inside your network while the gateway governs what reaches external models.

Shape AgentManager with us

Tell us about your agents and the models they use, and we'll show you where AgentManager fits.

Talk to Us About AgentManager