Guardian Agents at the AI Gateway: Wayfound is Now a One-Click Supervision Layer
AI agent teams have spent the past several years getting better at observing what their systems do. They can see model calls, latency, errors, token usage, and tool activity across increasingly complex workflows.
But observing an AI agent is not the same as supervising one.
A dashboard may show that a model responded in 1.8 seconds, called a retrieval tool, and consumed 2,300 tokens. It does not tell you whether the agent followed company policy, gave the customer the right answer, handled a handoff correctly, or achieved the business outcome it was designed to deliver.
Closing that gap has usually meant an integration project: add an SDK to every agent, wire up every framework, coordinate with every team that ships one.
There is now a much shorter path. If your LLM traffic already flows through an AI gateway, the gateway itself can become your supervision on-ramp. Now connecting an agent to a Wayfound Supervisor (called a “guardian agent” by Gartner) can be just a single click away.
Before: The Gateway Sees Everything… and Tells You Almost Nothing
Most organizations running AI at any scale have already put a gateway like LiteLLM in front of their model traffic. It is where they manage provider keys, control costs, enforce rate limits, and issue one virtual key per application.
That means the gateway already sits on top of the most complete record of AI activity in the company: every prompt, every response, every model, every application — including the internal tools, vendor apps, and legacy systems nobody is going to re-instrument.
For the business owner responsible for those agents, the important questions go beyond latency and cost. The concerns are critical to business goals and outcomes:
- Did the agent resolve the customer's request?
- Did it stay within approved guidelines?
- Did it use the correct knowledge?
- Did it escalate when it should have?
- Are recurring failures becoming more or less common?
- When company guidelines and agent directives conflicted, did the agent make the right choice?
The gateway has the raw material to answer all of that. Until now, there was nowhere to send it so that the process could actually answer these questions and actively direct the agent.
Now: Every Gateway Key Becomes a Supervisable Agent
Wayfound now connects directly to this layer. Point your AI gateway's (such as LiteLLM or other) telemetry at Wayfound (a configuration change on the proxy, not a code change in any application) and watch your trust in your agentic deployments increase.
Now every virtual key that sends traffic through the gateway shows up in your Wayfound settings automatically, with its alias, team, and traffic stats. Your gateway's key list becomes a live inventory of the AI applications running in your organization, including the ones you didn't know about.
From there, enabling supervision is just one click. Flip the Supervise toggle on a key and Wayfound:
- creates a Supervisor for that application,
- connects it, and
- backfills roughly the last week of that key's traffic
So your first analyzed sessions appear within minutes of connecting. No agent IDs to copy, no client code to change, no deployment. Most teams already issue one virtual key per application for cost tracking; if that's you, the key is the agent's identity, and key rotation is handled automatically.
The Supervisor acts as a Guardian Agent: an AI system responsible for overseeing another AI agent against defined goals, operating guidelines, and business expectations. Instead of merely showing what happened, it evaluates what the activity means, flagging guideline compliance issues, knowledge gaps, unsuccessful outcomes, problematic sessions, and trends across large volumes of interactions.
Gateway traffic arrives as spans and API calls; Wayfound reconstructs it into readable sessions interpreting raw data into user messages, assistant responses, model usage, token counts, latency, and errors, so that the work of the Agent can be actively supervised and improved.
One Small Field Unlocks Multi-Turn Supervision
The gateway path needs almost nothing from application teams. The one thing worth adding is a stable session_id, so multi-turn interactions group into a single session that can be evaluated holistically, not just as individual requests and responses:
{
"model": "...",
"metadata": { "session_id": "chat-123" },
"messages": [...]
}
- session_id merges every request in a conversation into one ordered session for review.
- Running multiple agents behind a single key? Add wayfound_agent_id to the metadata to route each request to its own Supervisor; it overrides the key's default agent.
That's the entire integration surface for a typical application.
Why the Gateway is the Right Place to Start
It covers what you can't instrument. Vendor applications, legacy systems, and tools owned by other teams all send their model traffic through the same proxy. The gateway path means CIOs, CISOs, CPOs, and other leaders can access and supervise their AI behavior without asking anyone to change their code.
It gives you an inventory, not just a pipeline. Because discovery is automatic, connecting the gateway answers a question many organizations struggle with: what AI is actually running here? Every key that carries traffic surfaces in one place, with the option to put a Guardian Agent to supervise everything that’s happening.
It scales governance the way gateways scaled cost control. Teams adopted gateways so that provider keys, spend, and rate limits could be managed centrally instead of app by app. Supervision follows the same logic: one connection point, consistent business and governance standards, every application evaluated against its own goals and guidelines in one place.
It's a starting point, not a ceiling. The gateway path provides every LLM call's messages, model, token usage, latency, and errors. When you want richer structure for a particular agent or multi-agent workflow, you can layer in framework instrumentation for just that application, and Wayfound merges the two views automatically, de-duplicating so the framework's record of each call is the source of truth and the gateway's fills in token and cost details.
Built on the Standards You Already Run
Under the hood, this all rides on OpenTelemetry — the same telemetry standard your infrastructure likely already speaks. LiteLLM's built-in otel callback is the gateway connection, and Wayfound also understands the semantic conventions used by agent-framework auto-instrumentors (OpenInference and the OpenTelemetry GenAI conventions) for teams that want the deeper, framework-level view of agents built with LangGraph, the Claude Agents SDK, the OpenAI Agents SDK, Workato, LlamaIndex, and others.
But the standard is the plumbing, not the point. You don't need to think about spans, exporters, or trace semantics to supervise the applications behind your gateway. You need a gateway that's already carrying the traffic and one toggle in Wayfound.
Observability Tells You What Happened. A Guardian Agent Tells You Whether it was Acceptable and How to Improve.
Wayfound's gateway integration does not need to replace the observability tools your engineering teams rely on. It uses the same infrastructure for a different purpose.
Observability helps teams understand whether the system is functioning. Guardian Agent supervision helps teams understand whether each agent is behaving correctly, meeting expectations, and producing the outcomes the business needs. When something goes wrong, the Supervisor can stop the agent, suggest improvements, and update the agent with new directives to ensure it improves over time.
That is the real before-and-after. Before, your AI gateway could tell you what every application spent. Now it can tell a Guardian Agent whether every application did its job.
Learn more at docs.wayfound.ai
See Wayfound's AI Gateway integration in action. Request a Demo.
-----------------------------------------------------------
Frequently asked questions
What is a Guardian Agent? A Guardian Agent is an AI system that oversees other AI agents. It evaluates their activity against defined goals, policies, guidelines, and expected outcomes, helping organizations identify performance problems, compliance risks, and knowledge gaps. Wayfound has been recognized by Gartner as a leading Guardian Agent platform in the Business Alignment & Outcome Optimization category.
How does gateway-based supervision work? Your AI gateway (ie. LiteLLM) sends its telemetry to Wayfound. Every virtual key seen in that traffic appears in Wayfound settings as a discoverable agent. Flipping the Supervise toggle on a key creates its Supervisor, connects it, and backfills roughly the last week of traffic, so analyzed sessions appear within minutes.
Do application teams need to change their code? No code changes are required for basic supervision — the gateway is the only integration point. Adding a session_id metadata field enables multi-turn session grouping, and a wayfound_agent_id field routes requests when multiple agents share one key.
Do I need the Wayfound SDK? Not for agents behind a supported gateway. The SDK and framework-level instrumentation remain available when you want deeper structure: tool calls, agent steps, and handoffs in multi-agent systems.
Can Wayfound supervise agents that aren't behind a gateway? Yes. Agents built with common frameworks (LangGraph, CrewAI, Claude Agents SDK, OpenAI Agents SDK, Workato, LlamaIndex, and others) can send their telemetry to Wayfound directly using standard instrumentation, and each agent is routed to its own Supervisor.
What is the difference between AI agent observability and AI agent supervision? Observability focuses on technical behavior such as latency, errors, model calls, and token usage as well as standard single-turn evals. Supervision evaluates whether the agent followed guidelines, handled the interaction correctly, and achieved the desired business outcome in the context of the business’ goals, policies, and culture. The Wayfound Supervisor also iteratively learns the organizational context in its memory and continues to add context to all your AI agents as they are working.