Platform · Runtime assurance

Most tools show how an agent was set up. Govern360 shows what it actually did.

Govern360 receives observed runtime evidence from the OpenTelemetry collector you already run. It is not a hop in your request path, adds no latency, and cannot stop a call at request time. Every product that sees runtime by sitting in the path is a proxy. This one is not.

Real-time here means seconds to minutes after your collector exports. It has never meant inline blocking, and this page will not pretend otherwise.

PRODUCT ARCHITECTURE / VISUAL EXPLAINER

Understand the underlying model

Distinguish configured safeguards from observed runtime behaviour.

  1. 01 / Source event Identify the relevant signal and its source.
  2. 02 / Decision logic Add access, ownership and operating context.
  3. 03 / Measured result Review the state and retain the evidence.
Request a product walkthrough
Conceptual Govern360 runtime assurance diagram: Configured, Decided, Observed, Evidence. Illustration, not live customer data.
Illustrative workflow — not a live scan or customer result.
Architecture diagram. A user or app calls an AI agent directly, labelled Direct request path, no proxy, no hop added; the agent reaches a model, tools and data. Separately, OpenTelemetry runtime telemetry flows from that path into the customer's own OpenTelemetry Collector, which exports observed runtime evidence to Govern360 seconds to minutes after export. Govern360 Runtime Assurance reports three states - Configured, Decided and Observed - across nine capabilities: intake, identity, release, findings, policy, spend, provider, SIEM and linkage. Govern360 sits beside the request path, never inside it.
The request goes straight from your app to the agent — Govern360 is not on that line. Your collector exports the traces; Govern360 receives them seconds to minutes later and reports what the agent was configured to do, what was decided, and what was actually observed.
Three states, not one

A decision is not an invocation

Nearly every product in the category draws these as one line on a graph. Govern360 did too, until it went looking: 359 observed edges across 180 agents turned out to be policy decisions, one of them a DENY rendered as “sends to ServiceNow, 3 calls”. The fix deleted 359 of the product’s own findings, because they were not findings.

How it works

Twelve capabilities, and what each one refuses to do

01

Runtime telemetry intake

Traces arrive in standard OpenTelemetry format from the collector you already run. Each connector carries its own token with expiry, rotation with overlap and a quota, at most two live at a time. Test connectors stay separate from production, so trial traffic never lands in production evidence.

Your collector exports. We receive. No hop is added.

02

Agent identity resolution

A trace is linked to a known agent only by an exact identifier. A name that merely looks right goes to a review queue for a person to confirm.

A wrong attribution is worse than none.

03

Release certification

Every production agent gets a tiered, signed approval tied to its model, purpose, tools and autonomy tier. Change any pinned part and the approval is marked changed rather than quietly staying valid. It also expires on a schedule.

An approval is a statement about a moment. Agents change weekly.

04

Runtime findings

Retired agents still running. Agents in production with no current certificate. Model drift. Tools called that were never approved. Data reaching destinations outside the allow list. Checked hourly, raised into the alert feed you already have.

Evidence, not enforcement. A finding never blocks anything.

05

Policy dry runs and replay

Ask the live decision engine what it would decide, with nothing enforced. Replay the last 30 days of observed tool calls against today’s rules to see what a proposed change would actually have done.

A dry run is never a decision. It never enters the decision log.

06

Runtime spend reconciliation

Each model call is marked confirmed by both sources, seen only at runtime, or seen only in spend records. The middle column is the interesting one: it ran, and it is in nobody’s bill.

Counting runtime-only usage is opt-in, and no closed month moves.

07

Provider posture

For each AI provider: the connection, keys, projects, admin log, spend and runtime activity, side by side. It writes nothing, calls no provider and reads no credential.

Gaps are named in words. Not visible, never “no keys”.

08

SIEM forwarding and the evidence pack

Runtime findings and replay summaries can be forwarded to your SIEM through the existing outbound channel, off by default. Auditors get an evidence pack with a SHA-256 checksum list.

Outbound only. Off until you turn it on.

09

Linkage maps everywhere

Every runtime surface draws the map rather than a table. Lines move only where something was observed, problems blink, a click opens the detail drawer, and Next issue steps through them.

Moving lines only where observed. Never for a decision.

10

MCP tool drift detection

An approved tool does not always stay the same. When an MCP tool’s definition changes after a human approved it — the rug pull, where a trusted tool quietly becomes something else — it is raised as a drift signal in the MCP Gateway and carried into the runtime verdict, so a changed-since-approved tool can be caught at the moment it is used rather than at the next review.

Drift against the approved fingerprint. Never against a guess at intent.

11

Runtime AI authorization

On Enterprise Claude surfaces the platform’s own inference hooks call Govern360 before a prompt runs and, with tool-call validation on, before a tool executes. Govern360 answers from the signed authorization credential it already holds for that agent, evaluating the tool, its side-effect class, the egress allow list, tool-definition drift and sensitive-data detectors against that agent’s contract. No proxy, no endpoint agent, nothing of ours in the path.

Allow or deny only. It never redacts, and it ships dormant.

12

Claude Platform telemetry

Anthropic’s OpenTelemetry export is protobuf, gzip-compressed and not configurable, so it cannot be pointed straight at a JSON intake. A guided setup routes it through a standard OpenTelemetry Collector into Govern360, generating the exact collector configuration and destination steps for you. Metadata only, the same as every other runtime source.

The works-today path, stated as one. Native direct ingest is roadmap.

Questions

Runtime assurance, answered

Does Govern360 sit inline in the request path?

No. Runtime evidence reaches Govern360 because the customer’s own OpenTelemetry collector exports traces to it. Govern360 receives those traces; it is not a hop the request passes through and adds no latency, so telemetry on its own never stops a call. Where a decision is wanted before an action runs, the arrangement is the same one used everywhere else — Govern360 decides and the plane enforces. On Enterprise Claude surfaces that plane is the platform’s own inference checkpoint, which asks Govern360 for a verdict and acts on the answer. Every product that sees runtime by sitting in the path is a proxy you have to defend a latency budget for. This one is still not.

What does Govern360 mean by real-time?

Seconds to minutes after the source exports, and it never means Govern360 became a hop in the path. The figure depends on the customer’s collector export interval, not on anything Govern360 controls, so the honest description is that telemetry evidence lands shortly after the span is exported rather than at the moment of the call. A runtime verdict is a different mechanism and should not be read off this number: there the platform’s checkpoint asks Govern360 before the action runs.

What is the difference between what an agent is configured to do and what it actually did?

Configured means a source configuration permits or requires it — read from the agent’s setup, which it may never have exercised. Observed means a source event proves the call completed. Govern360 reports both, and never merges them: a policy decision such as a DENY is a permission or a refusal, not an observed action, and is never drawn as one.

How does runtime telemetry reach Govern360?

Traces arrive in standard OpenTelemetry format from the collector the customer already runs. Each connector holds its own token with an expiry, rotation with overlap and a quota, and at most two live tokens at a time. Test connectors are kept separate from production so trial traffic never lands in production evidence.

How is a trace matched to a known agent?

Only by an exact identifier. A name that merely looks like a match is not accepted — it goes to a review queue for a person to confirm. Guessing here would attach real runtime behaviour to the wrong agent, which is worse than leaving it unattributed.

What is Release Certification?

A tiered, signed approval for a production agent, tied to its model, purpose, tools and autonomy tier. If any of those pinned parts change, the approval is marked changed rather than silently remaining valid, and it expires on a schedule. An approval is a statement about a moment; agents change weekly, so the certificate has to know when it has been outrun.

What do runtime findings actually catch?

Agents that were retired but are still running, agents in production with no current certificate, model drift away from what was approved, tools being called that are not on the approved list, and data going to destinations outside the allow list. They are checked hourly and raised into the existing alert feed rather than a second parallel pipeline.

Does a runtime finding block or stop an agent?

No. A finding is evidence, not enforcement. It never blocks, never stops an agent and never changes a certificate. Where an enabled policy blocks a tool and the source reports the call completed anyway, the finding says the block did not hold — and it says that only when the source actually reported completion, never on suspicion.

How does runtime authorization work with no proxy and no agent?

The platform holds the checkpoint. On Enterprise Claude surfaces the platform’s inference hooks call Govern360 before a prompt runs and, with tool-call validation enabled, before a tool executes; Govern360 answers from the signed authorization credential it already holds for that agent. Nothing is deployed into agents, browsers or endpoints, and Govern360 is still not a hop the request passes through — it is asked, out of band, by the platform that is.

Is runtime authorization enforcing anything today?

No. It ships dormant and shadow-first: it logs the verdict it would have returned, at a staged rollout percentage, with a kill switch, and nothing is enforced until an owner turns it on. It is also fail-open by construction — if Govern360 cannot answer, the action proceeds rather than the platform stalling. A control that takes the business down when the governance layer has a bad day is not a control anyone keeps.

What does a runtime verdict evaluate, and what is stored?

Per tool call: tool allow or deny, the side-effect class (read, write, send, delete, deploy), the egress allow list, tool-definition drift since approval, and sensitive-data detectors, all against that agent’s contract. Govern360 reads the transcript and tool calls only to decide and stores only the verdict and its metadata, never prompt, response or tool-argument content. Every enforced block is written to the sealed, hash-chained audit record.

What is runtime authorization not?

Allow or deny only — it does not redact content. Image-only requests are not inspected. It covers Enterprise Claude surfaces rather than Bedrock, Vertex or the raw API, and some native tool calls are not covered. It is one control in a layered programme, not a sole control, and saying otherwise would be the overclaim you discover at the worst moment.

What is MCP tool drift, and why does it matter?

An approved tool does not always stay the same. When an MCP tool’s definition changes after a human approved it — the rug pull, where a trusted tool quietly becomes something else — it is raised as a drift signal in the MCP Gateway and carried into the runtime verdict, so a changed-since-approved tool can be caught at the moment it is used rather than at the next review. Drift is measured against the approved fingerprint, never against a guess at intent.

How do I connect Claude Platform telemetry to Govern360?

Through a standard OpenTelemetry (OTel) Collector. Anthropic’s OpenTelemetry export is protobuf, gzip-compressed and not configurable, so it cannot be pointed straight at a JSON intake. Govern360 gives you a guided setup that generates the exact collector configuration and destination steps, and the collector translates the export into the JSON intake. Metadata only, the same as every other runtime source. Native direct ingest is on the roadmap — this is the path that works today.

Can we test a policy without enforcing it?

Yes. A dry run asks the live decision engine what it would decide, with nothing enforced. A dry run is never a decision: it reads not enforced, never enters the decision log, and never opens an approval request. You can also replay the last 30 days of observed tool calls against today’s rules to see what a proposed change would have done.

How does runtime evidence change what we are billed?

By itself, not at all. Each model call is marked as confirmed by both sources, seen only at runtime, or seen only in spend records. The interesting column is the middle one — it ran, and it is in nobody’s bill. Counting runtime-only usage is opt-in per organisation and counts from the moment it is switched on, so no closed month ever moves.

What is provider posture?

A single read, per AI provider, putting the connection, keys, projects, admin log, spend and runtime activity side by side. It writes nothing, calls no provider and reads no credential. Gaps are named in words: a provider with no connection reads not visible, never no keys; an admin log that could not be read says so with the reason.

Can runtime findings go to our SIEM?

Yes, outbound only and off by default, through the existing SIEM forwarding rather than a new channel. For auditors there is a runtime assurance evidence pack with a SHA-256 checksum list, so what you hand over can be shown not to have changed since it was generated.

Does runtime telemetry mean you now store our prompts?

No. The capture mode permits metadata only and nothing else is accepted. Prompt text and model responses are never stored. What arrives is span metadata — which model, which tool, which destination, how many tokens — not what anyone typed.

Can you see which employee triggered an agent?

Not by name on default screens. The initiator is stored as an HMAC under a per-organisation key held inside the database, together with the type of initiator. Runtime findings are about agents, not scores of people.

What do the runtime screens show before we connect telemetry?

They say no runtime telemetry is connected. They never show zero runs, because zero is a measurement and this is an absence of one. If a connector exists but nothing arrived in the window, it says no events were observed in connected telemetry during the selected period — which is a different statement, and the difference matters.

What does Govern360 see at runtime?

What the agent actually did — from traces exported by the customer’s own telemetry collector, without sitting in the request path. That covers runtime intake, agent identity resolution, release certification, runtime findings, policy dry runs and replay, runtime spend reconciliation, provider posture, SIEM forwarding with an evidence pack, and linkage maps on every runtime surface.

What this page does not claim. Govern360 is not in the request path and cannot block a call at request time. Runtime findings are evidence and never enforce anything. Prompt text and model responses are never stored — metadata only. The initiator is pseudonymous. Where a source cannot be read, screens say so with the reason rather than showing zero. Figures cited here come from a connected demonstration tenant, not a customer estate.

More runtime questions, answered · Why Govern360 · The data-to-action graph · Control AI spend · All answers