Agent Operations Platform
Audit
Every governed action emits an immutable event with the context that produced it: tool call, model call, lifecycle change, policy decision. One vocabulary covers all three control points, so one query reconstructs an incident wherever it happened.
Workflow
Capture an action.Stream it everywhere.
The same loop for every event. Captured at the source, enriched with context, streamed live, exported to wherever your SOC already works.
- 01
Audit
Every governed action emits an immutable event at the source — Registry, Gateway, or Broker.
- 02
Observe
Query, tail, and filter from the UI, CLI, or API. Full caller chain on every event.
- 03
Stream
Subscribe to the live event stream for real-time monitoring, alerts, and downstream automation.
- 04
Export
Project events to your SIEM, lakehouse, or compliance evidence store. Dome remains the system of record.
Consistent operations
One vocabulary, three control points
One event model turns three enforcement points into one accountable system. Registry, Gateway and Broker each emit their own classes, in the same shape: same caller chain, same decision trace, same export targets.
Each control point emits a typed event at the moment of decision or transition.
Caller chain, scope, policy version, and request digest attached automatically.
Live tail, query, and subscription APIs serve the events in real time.
Projected to SIEM, lakehouse, or evidence store. Dome stays the system of record.
Unified vocabulary
agent.* from the Registry, tool.* from the Gateway, llm.* from the Broker, policy.* across all three. One schema, one query language, one story per incident.
Caller chain
Every event names the agent, the person it acted for, the parent agent if chained, and the rule version behind the decision.
Multiple sinks
Query from the UI or the CLI, subscribe through the API, or export to Splunk, Snowflake, S3, or whichever sink your team already runs.
Options
What audit lets you do
What you can ask of the record, and what it guarantees when you do.
Query and live tail
Filter by event type, class, result, agent, resource, actor, or originating surface, and by fields inside the payload. The same filters narrow a live tail.
Activity correlation
Events chain into activities, so a session, a turn or a run reconstructs in order. One identifier walks the whole thing.
Attribution, down to the surface
Every event records what produced it — SDK, CLI, API, MCP, dashboard, or the data plane. When a filter redacts or blocks, each action is attributed to the filter that took it.
Export for your SIEM
OCSF for Splunk, Sentinel or Chronicle, CEF for ArcSight or QRadar, CSV and JSON for a lakehouse. Reads and exports report their own completeness, and one that cannot prove it is refused.
A published event catalog
Enumerate every event type and its schema programmatically. An integration builds against a declared contract, not a sample payload.
SDK-emitted events
An agent can emit its own events from inside the process — decisions, errors, custom signals. They flow through the same pipeline, with the same caller chain attached.
In practice
From query to export
Query it from the UI, the CLI, or the API, and export the same events wherever evidence has to live. Every event carries the whole caller chain: this agent, for this person, calling this tool, with these arguments, against this rule version.
# What did this agent do recently?
dome audit query --agent-id support-triage --since 24h
# Live tail
dome audit stream --agent-id support-triage
# Everything that was denied
dome audit query \
--results denied \
--since 24h# OCSF for Splunk, Sentinel, or Chronicle
dome audit export \
--format ocsf \
--since 24h > events.json
# CEF for ArcSight, QRadar, or syslog
dome audit export --format cef --since 24h
# CSV for a lakehouse or object store
dome audit export --format csv --since 24h > events.csv{
"event": "tool.call",
"ts": "2026-06-02T14:03:22.418Z",
"agent": {
"id": "agt_3mw7kp",
"name": "support-triage",
"workspace": "support"
},
"actas": {
"subject": "jane.smith@example.com",
"verified_by": "oidc"
},
"tool": "zendesk/update_ticket",
"arguments_digest": "sha256:9b...",
"decision": "allow",
"policy": {
"version": "v23",
"rule": "workspace/support/zendesk.cedar#L12"
}
}
Registry
Gateway
Broker
ManagementQuestions
Common questions about agent audit
What does Dome record?
Every governed action — tool call, model call, lifecycle change, policy decision — emits an immutable event with the full context that produced it.
What is in an audit event?
The full caller chain: this agent, on behalf of this user, calling this tool, with these arguments, against this rule version, with this outcome. That is the level of detail compliance and incident response need.
Can one query cover tools, models, and agent lifecycle?
Yes. The event vocabulary is uniform across the three control points, so a single query can reconstruct an incident regardless of which one it happened at.
How is audit different from eventing?
Audit is the record and eventing is the reflex. Audit is complete, immutable, and retained for evidence. Eventing projects selected facts onto a versioned public contract for delivery elsewhere, keeping the internal schema out of the API.