System Context
Sentry AI Agent Management Platform Architecture Diagram
Read it top to bottom: the people who run the business at the top, the Sentry Agent Management Platform in the middle, the agents and systems below. Every agent, whatever harness it runs in, is registered into the platform once, then monitored, cost-tracked and governed. The four arrows are the integration tiers, showing how deeply each one connects.
Operators
Leadership · Business teams · IT & Ops
Sentry Agent Management Platform
The control plane for your whole agent estate, delivered neutrally across every vendor.
Registers, monitors, governs and optimizes every AI agent the business runs, from one place, without building or replacing any of them.
Your agents do the work. Sentry stands watch.
Tier 1 · Inline
Sentry sits in the traffic path (gateway + MCP).
Tier 2 · Native telemetry
The harness streams OpenTelemetry to the platform.
Tier 3 · API pull
Poll the vendor's admin or audit API on a schedule.
Tier 4 · Registered-only
Register, set boundaries, govern by policy.
Bespoke / in-house agents
Custom builds, open-weight & self-hosted models
You own and host them, so they run behind the Sentry gateway and MCP layer: full transcripts, tokens, tool calls and live enforcement.
Hyperscaler agents
Azure AI Foundry, AWS AgentCore, Google Vertex
Cloud agent runtimes that emit OpenTelemetry natively. One config change streams every trace, token and tool call to the platform. No rebuild.
Frontier model agents
Claude, OpenAI, Gemini, Microsoft Copilot
Agents and assistants from the big model vendors. Their SDKs stream telemetry (Tier 2), and their org usage and compliance APIs are pulled for cost and transcripts (Tier 3).
Embedded agents
Salesforce Agentforce, CrowdStrike Charlotte, ServiceNow
Vendor agents baked into SaaS. Deep session data is pulled where the vendor exposes it (Tier 3); true black boxes are registered and governed by policy (Tier 4).
The register is built and the fast tiers connect: your own agents route through the gateway (Tier 1) and cloud runtimes stream the telemetry they already emit (Tier 2). You see live conversations and real cost the same day. First proof on the wall: an agent whose vendor reported $0.00, costed from the tokens the platform watched go past.
The closed platforms come in on their admin APIs (Tier 3), the black boxes are registered and bounded (Tier 4), the first approval gate goes live, and the cost baseline is set across the whole estate.
What it is, and what it isn't. Sentry AI is the control plane for your agents. It sits neutrally on top of everything you run and registers, monitors, governs and optimizes every agent from one place, whichever vendor or harness they live in. It is not an agent framework or builder, not a runtime or host, not agent memory or a RAG store, and not a model. It holds none of your context or data, and it does not rip and replace: it sits on top of your existing stack.
The top three benefits. First, you finally see and cut AI spend: one number for what AI costs across every vendor, then optimised down, often by 20 to 40%. Second, you govern every agent safely, with guardrails, human approvals, a full audit trail and a kill-switch. Third, you prove ROI and scale with confidence, from one source of truth, deploying more agents without ever losing control.
Where the line sits. Sentry runs the control plane end to end: the gateway, the telemetry collectors and vendor adapters, the registry and tier classification, the cost engine, the approval queue and the optimisation loop. Your team keeps what makes the agents yours: the agents and harnesses, the prompts and logic, the data sources and vendor relationships, and the final say on every approval. The intersection is deliberately thin: one registration per agent, plus the telemetry it already emits.