> ## Documentation Index > Fetch the complete documentation index at: https://docs.egisai.co/llms.txt > Use this file to discover all available pages before exploring further. # Coverage > The five places an AI agent can exist, the control EgisAI has on each, and how to tell which ones you have deployed. Governance is only as good as its worst blind spot. A workspace with a perfect policy set and one uninstrumented Go service has a perfect policy set and a leak. So EgisAI models coverage explicitly: five surfaces, one control on each, and a page that tells you which ones are actually in place. The **Coverage** tab under Protection reports it, and the report you can hand an auditor is generated from the same numbers. ## The five surfaces | Surface | What lives there | The control | | -------------------------------- | ------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------- | | **Your code, Python** | Agents you wrote and can `pip install` into | [The SDK](/installation) — in-process, before the call leaves | | **Your code, other languages** | Go, Java, Rust, TypeScript services calling a model API | [The Gateway](/gateway/overview), or the [egress node](/coverage/egress-node) where you can't change a base URL | | **Vendor-hosted agents** | Bots running on someone else's infrastructure | [Integrations](/integrations/agent-frameworks) — their calls never traverse anything we run, so the control is cutting their access | | **People using AI in a browser** | Someone pasting a customer list into a chat product | [The browser extension](/coverage/browser-extension) | | **Proof** | Evidence the four above are complete | The Coverage tab and its exported report | ## Why the surfaces are listed even when empty A coverage report that silently omits the surface it doesn't cover is worse than no report. It reads as complete, and the one number an auditor would have asked about is the one that isn't there. So every surface is returned on every read, including the ones with nothing deployed. A surface with no control says so, in words, with the name of the thing that would fill it. ## Honesty about what each control did Every audit row carries two fields that exist purely so a wall of verdicts can be read correctly months later: * **`enforcement_status`** — `enforced` means we were in the path and the call went the way the row says. `advisory` means we saw it and did not gate it. An egress node in observe mode and the browser extension only ever write `advisory`, because neither stopped anything. * **`inspection`** — `full` means policy ran on a real body. `metadata_only` means we saw a connection and never decrypted it, so any verdict on that row came from the destination or the volume, never the content. `blind` means it was reconstructed after the fact from a vendor's audit log. See the [enforcement matrix](/concepts/enforcement-matrix) for the per-framework version of the same question. ## What's next A proxy you run. Two environment variables on a workload and every model call it makes is inventoried — in any language, with no code change. Which AI sites your team uses, and optionally what kinds of sensitive data get typed into them. Point a base URL at EgisAI and every call through it is governed inline. Where the SDK prevents versus where it can only record.