Not every agent in your estate is one your engineers built. Teams ship Copilot Studio bots and Foundry agents without a line of deployable code — which means there's nowhere to add an SDK, and until now, no way onto the tape. The Wytness Collector closes that gap: a container you run that turns the Microsoft telemetry you already have into signed, PII-tokenised audit events. No code changes, anywhere.
The estate you actually have
The uncomfortable census in most organisations: for every agent an engineering team built and can instrument, there are several that arrived without a repository — a Copilot Studio bot the operations team assembled, a Foundry agent from a pilot that quietly became production. Audit scope doesn't care which kind acted on the customer record; a governance story that only covers the agents you built is a story with a hole where the auditor will look first. The Collector exists for the other kind.
How it works: poll, map, sign, ship
- Poll. The Collector reads new Microsoft Foundry tracing and Copilot Studio telemetry rows from Log Analytics on a cadence, resuming from a persisted cursor — no agent modification, just the logs your platforms already emit. The cursor means restarts and redeploys pick up exactly where they left off: no gaps, no re-reads.
- Map. Each row becomes a Wytness event with a deterministic ID, so re-reads and replays can never create duplicates — the same source row always maps to the same event, and the platform collapses any repeats.
- Sign and tokenise. Inside the container, each event is signed with your Ed25519 key and its PII fields tokenised — the same customer-held-key crypto the SDK uses. Your keys and your source credentials never leave the container you run; from the platform's perspective, a collector looks like any other customer environment holding its own secrets.
- Ship. Signed envelopes POST to ingest. If delivery fails, events dead-letter to the container's volume and replay automatically when the path recovers; duplicates collapse server-side. An outage delays the record — it doesn't delete it. In a live drill we ran during development, events captured against a dead endpoint were retained in full and replayed to an exact match once the endpoint returned.
The failure mode nobody watches for
With log-based capture, the scary scenario isn't an error — it's silence. A source that quietly stops producing looks identical to a source with nothing to say, and months of missing evidence can hide in that ambiguity: a permission expires, an export rule changes, a workspace moves, and every dashboard stays green while the record quietly stops. So the Collectors page tracks every source's poll health per instance, and a source that goes silent is flagged rather than missed. "No events" becomes a visible state you can investigate, not an absence you discover during an audit — which is the difference between a configuration hiccup and an evidence gap with your name on it.
An honest label, not a hand-wave
Collected events carry a genuinely different guarantee from SDK events, and we label it instead of blurring it. An SDK event is attested by the agent's own environment at the moment of action. A collected event proves the log content is untouched since pickup — real, valuable, and deliberately ranked below the SDK's guarantee. Every event shows which tier it carries, in the dashboard and in every Evidence Pack, so an auditor always knows which claim they're standing on. The alternative — pretending both capture paths prove the same thing — would be exactly the kind of claim this product exists to make impossible.
Where it fits
Use the SDK for agents you build; use the Collector for agents you merely operate. Most estates need both, and the dashboard treats them as one fleet — same Events page, same anomaly rules, same Evidence Packs, with the capture tier visible on every row. The gap it closes is the audit's classic blind spot: the low-code agents nobody instrumented because there was nothing to instrument. The full setup walk-through lives in the Collector docs. Questions about your sources? Ask us.