Skip to Content
Audit & Traceability

Agent Trail

Find who did what and when, with every tool call, model call, agent session, and setting change in one record

When an agent does something unexpected, a call is blocked, or someone asks who changed a rule, you need the record behind it. Agent Trail records every tool call, model call, and agent session that passes through Agent Authority, and every change to its settings. Each entry names who acted, what they acted on, and when.

Audit log showing tool execution events
Every action is logged with the actor, resource, and outcome

What's recorded

Each row lists events by the name the Action filter shows.

GroupActionWhat it records
Tool callsTool ExecutedA call that passed the rules and ran: the tool and the MCP server
Tool callsTool FailedA call that was blocked or failed, and why: blocked by a rule, refused by a usage limit, or an error
Tool callsTools SearchedEach tool search and the tools it returned, but not the query
Tool callsApproval Requested, Approval Granted, Approval Declined, Approval ExpiredA call that needed approval, and what became of it
Model callsModel CallEach model call through the gateway, including calls a rule denied
AgentsAgent ConnectedEach MCP session an agent opens, with the servers it could reach, was blocked from, or couldn't reach
AgentsAgent Created, Agent Updated, Agent Revoked, Agent DeletedAn agent added, changed, revoked, or removed
AgentsAgent TaggedA tag assigned to an agent
MCP servers and toolsMCP Created, MCP Updated, MCP DeletedA server added, changed, or removed
MCP servers and toolsTools SyncedEach time the gateway finds new or changed tools on a server, and the first time it reads a server's tools
MCP servers and toolsTools ApprovedNew or changed tools that were held back, approved by an administrator
MCP servers and toolsMCP Tagged, Tool TaggedA tag assigned to a server or a tool
Sign-ins and accessConnection Created, Connection Removed, Connection Revoked, Connection ExpiredA person's sign-in to a server that acts on their behalf
Sign-ins and accessMCP Request Created, MCP Request Withdrawn, MCP Request DismissedA person's request for access to a server
RulesPolicy Set, Policy Updated, Policy RemovedAn access rule added, changed, moved, or removed
RulesDLP Rule Created, DLP Rule Updated, DLP Rule DeletedA response filter added, changed, or removed
RulesRate Limit Created, Rate Limit Updated, Rate Limit DeletedA usage limit rule added, changed, or removed
RulesTag Created, Tag Updated, Tag DeletedA tag added, changed, or removed
RulesAudit Policy Created, Audit Policy Updated, Audit Policy DeletedAn audit policy added, changed, or removed
AdministrationAPI Key Created, API Key Rotated, API Key RevokedAn API key added, rotated, or revoked
AdministrationRole Created, Role Updated, Role Deleted, User Roles ChangedA role added, changed, or removed, or a person's roles changed
AdministrationUser Deleted, Group DeletedA person or a group removed
AdministrationGateway Settings UpdatedA change to the gateway's settings
AdministrationWebhook Created, Webhook Updated, Webhook Deleted, Webhook DisabledAn event forwarding webhook added, changed, removed, or turned off
AdministrationIngest Key Created, Ingest Key RevokedA telemetry ingest key added or revoked
FindingsAnomaly Detected, Anomaly Escalated, Anomaly Resolved, Anomaly Status ChangedAn anomaly found, escalated, resolved, or triaged
FindingsFinding Status ChangedA status change on a non-human identity finding
CostsSeat Price Set, Seat Price Ended, Seat Price DeletedA seat price set, ended, or removed; a new rate names the rate it replaced
CostsSeat Assigned, Seat Assignment Ended, Seat Assignment RemovedA plan assigned to a person, ended, or removed

Approvals, access requests, held-back tools, model calls, audit policies, and non-human identity findings appear only when SecureAuth has turned those features on for your organization.

What each entry shows

  • Timestamp – when the event happened
  • Actor – the person or agent that acted
  • Action – the event, by the name the Action filter shows
  • Resource – the MCP server, agent, rule, or other item involved
  • Details – what the action did, such as the tool name, the error, or the fields that changed

When the event records them, an entry also shows the IP address and the client it came from.

Entries with more to show have a chevron: agent connections, tool calls, tool searches that returned tools, model calls with a redaction or a recorded prompt, Tools Synced entries with more tools than fit on the row, and changes that list fields before and after. Select it to expand the entry. Anomaly entries link to the anomaly instead.

See which rule decided a call

Expand a Tool Executed or Tool Failed entry to see the tool, the MCP server, the agent, and the decision.

  • Allowed – the call ran, and the entry names the rule that allowed it.
  • Blocked – the entry names the rule that blocked the call, and the reason. A call that matched no rule is blocked by default, so its entry names no rule.
  • Refused by a usage limit – the entry names the limit. A limit set to log only lets the call run and still names the limit on the entry.
An expanded tool call entry showing the rule that allowed it and the response filters that redacted it
An allowed call names the rule that allowed it

Expand an Agent Connected entry to see what the agent could use when it opened the session. Each server it could reach shows how many of its tools were allowed, conditional, or denied. A conditional tool runs only when an argument-level condition passes on that call, or only after someone approves it. Expand a server to see its tools grouped under the rule that decided them. Tools that matched no rule appear under No matching rule · default deny. A blocked server names the rule that blocked it, or says no rule grants access. A server the agent couldn't reach shows the reason. Each rule name links to the rule in Agent Access.

An expanded agent connection entry with tools grouped under their deciding rules
Each tool sits under the rule that decided it, and a blocked server names the rule that blocked it

See what was redacted

A tool call whose response was redacted carries a Redacted ×N badge, where N is the total number of spans replaced. Hover over the badge or move focus to it to see how many spans each response filter replaced. Expand the entry to see the same counts, with each filter name linked to it in Data Protection. Agent Trail stores the filter names and counts, never the redacted text.

Model calls carry the same badge when a Data Protection rule redacted the prompt or the response.

A filter in log-only mode adds a Would redact ×N badge instead, and its counts appear under DLP log-only matches when you expand the entry. The agent gets the content unchanged. End users never see log-only matches.

What end users see

End users see a narrower version of this feed. Their Activity page shows only their own events, and only these: their tool calls, model calls, tool searches, approvals, access requests, agent events, and sign-ins to servers. Detail payloads are stripped, so they see which tools an agent could use but never which rule decided it. On an agent connection event, blocked servers, denied tools, and why a server couldn't be reached appear only in Agent Trail. Tool rows stay single-line, and the redaction badge reads a generic Content redacted without naming the rules or counts. Administrative changes, and anyone else's activity, are visible only in Agent Trail.

How long Agent Trail keeps events

Agent Trail keeps three months of history, then deletes it. To keep events longer, or to investigate them in your SIEM, forward them as they happen with Event forwarding.

Filtering

Use the filter dropdowns at the top of Agent Trail to narrow the feed:

  • Action – filter by event, using the names in What's recorded
  • Agent instance – see activity for a specific connected client
  • Agent – roll up across every instance of one agent (for example, all Claude Code clients)
  • User – events attributed to a specific gateway user
  • Group – events from the current members of a user group
  • MCP – see events related to a specific MCP server
  • Agent tag – events from any agent carrying a tag (for example, all prod agents)
  • Resource tag – events on any MCP server or tool carrying a tag
  • Autonomous – choose Only for events by autonomous agent instances, or Exclude for events by user-bound ones

Agent Trail lists the newest events first. Scroll down and click Load more to see older ones.

Excluding values

The Action, Agent instance, Agent, User, Group, and MCP filters can also hide events. Click a value to include it. Click again to exclude it, and once more to clear it. The include and exclude buttons on each row set a value directly.

An exclusion keeps events that have no value for that field. Excluding an agent still shows events that name no agent, such as policy changes. Excluding a group hides events from its current members. The Agent tag and Resource tag filters only include.

The Autonomous filter works differently. Both Only and Exclude keep just the events an agent performed, so policy changes and other events with no agent drop out. Events recorded before SecureAuth tracked this flag match neither choice.

Find what you need

To answerFilter by
Why a call was blockedAction set to Tool Failed, with Agent or User, then expand the entry
Which tools an agent got when it connectedAction set to Agent Connected, with Agent; the oldest entry is its first session
Who changed an access ruleAction set to Policy Set, Policy Updated, or Policy Removed
Whether a response filter is workingAction set to Tool Executed, then look for the Redacted ×N badge
Every tool call to one MCP serverMCP
Only setting changes, without the tool callsAction, with Tool Executed excluded

Moving a rule records Policy Updated with its new position, whether you drag it or set Place at when you edit it.

Why an agent can't reach an MCP server

An Agent Connected entry lists each server the agent couldn't reach, with the reason. The MCP filter doesn't find these entries, so filter by Action and User instead.

The entry saysWhat it means
its server could not be reachedThe gateway couldn't connect to the MCP server
the user's sign-in could not be refreshedThe person's own sign-in to the server couldn't be renewed
it needs a user, and this agent belongs to the organizationThe server acts for a person, and an organization-owned agent has none
enterprise-managed authorization failedEnterprise-managed authorization failed; the entry names where it failed and the error code

A server that kept its last known tools is listed once, as allowed, with the reason under it. A server made of several services lists each service that failed with its own reason, even when the rest connected. To see which servers are failing and for whom, see MCP server health.

Next steps

You can now find who did what in Agent Trail, see which rule decided a call, and trace why an agent couldn't reach a server. Next, forward events to your SIEM so you keep them past three months.

On this page