A Receipt for Every Action: Audit Logs for AI Agents
Here's an illustrative review. Your compliance lead asks three questions before the quarter closes. Who changed the production AWS connection last month? Which agent actions ran against it afterwards? Who approved them?
With people, you'd answer from tickets, chat threads and memory. With an agent in the loop, memory isn't an answer. The agent doesn't remember last month, and the person who approved a write at 3 a.m. may not either.
We built Frontier Investigation Agents to leave a record of every action, so a question like that has one place to go.
Why the usual approach breaks
Most teams bolt an AI tool onto systems that already log. The cloud provider logs API calls. The Git host logs merges. So the theory goes that you can reconstruct what the agent did from the systems it touched.
You can reconstruct what happened. You can't reconstruct why, or who allowed it. Your provider's log shows a call from a role. It doesn't show which conversation asked for it, which person approved it, or whether a person was involved at all.
An agent adds a third kind of actor next to people and the platform. The log has to name it as one.
Three kinds of actor
Audit Logs record user, agent and system actions across your organization in a single timeline at Admin Settings → Audit Logs.
01User
A person
Someone in your organization, with the IP address the action came from.
02Agent
An AI agent
An agent acting in a workspace, labelled as itself.
03System
The platform
The platform itself, separate from any person or agent.
Every entry carries an actor type. User is a person in your organization. Agent is an AI agent acting in a workspace. System is the platform itself. Filter to Agent and you see every action an agent took, across every workspace in the organization.
What gets recorded
Actions land in the timeline as they happen: connection changes, approval decisions, credential updates, billing events, user management, and activity from Optimize, Review, Cyber and Resolve. Each entry carries a category matching the surface it came from, such as Connection, Approval, Security or Billing.
A row shows the timestamp, actor, category, event, severity, a one-line summary and the workspace. The detail view adds the resource acted on (type, name and ID), a link to the related Resolve incident when one exists, and the IP address a user action came from.
Each event also carries an outcome: Success, Failure or Pending. Severity runs Low, Medium, High and Critical.
Credentials are logged with care. Every credential create, update and delete is recorded under the Security category, by key name and never by value. The log is a record of the secret's lifecycle, not a second copy of it. See The Agent Never Holds Your Keys.
Entries can't be edited
Entries cannot be edited or deleted from the app. That's the property that makes the timeline usable in a security or compliance review: no one can tidy it up from the app after the fact.
Viewing it is limited too. Only organization Owner and Admin roles can open Audit Logs.
Running a review
01Filter
Scope the question
Actor type, category, severity, date range or free text.
02Read tiles
Check the totals
Total Events, Success, Failures and Agent Actions for the filter.
03Open a row
See the detail
Resource, related incident and the IP address of a user action.
04Export
Keep the record
audit-trail.csv with your filters applied, before the 90 days run out.
Every filter narrows the timeline, the summary tiles and the export together. You can search free text, and filter by actor type, category, severity and date range.
Above the table, tiles show Total Events, Success, Failures and Agent Actions for the current filter. That's the quick read on whether anything needs attention before you scroll.
Export CSV downloads audit-trail.csv with your current filters applied. For the compliance lead's three questions: filter to the Connection category and last month's dates. Then set the actor type to Agent and search for the connection's name. Then switch to the Approval category. Export each view, and the scoped reports are the answer.
Where the rest of the record lives
The audit log answers "who did what". It isn't the only record, and it isn't meant to replace the others.
01Conversation
What the agent did
- Tool calls and outcomesStay in the conversation after approval
- Autonomous runsEach keeps its conversation
Step by step
02Audit Logs
Who did what
- User, Agent, SystemAcross every workspace
- Cannot be edited or deletedFrom the app
For security and compliance
03Activity
What agents reported
- Searchable inboxFindings, reviews, incidents, conversations
- Filter by moduleOr unread only
For finding past work
The conversation keeps the detail. When an approval pauses a tool call, the call and its outcome stay in the conversation. Every autonomous run keeps its conversation too, so you can open a scheduled run and read exactly what the agent did. Sessions from the Agent CLI are mirrored into the workspace as read-only conversations.
The investigation keeps its own history. Re-investigating a Resolve incident updates the same record, and a changed conclusion is recorded as a root-cause revision rather than a second report. See Inside an Investigation.
Activity is the searchable inbox of everything agents have reported across modules. It's for finding a past finding or review, not for proving who changed what.
The design decisions
Agents are first-class actors. An agent action is labelled as one, not folded into the user who started the chat or the platform that ran it.
One timeline for the organization. A review shouldn't need a login to every workspace. The log spans all of them.
Approvals are events. A write that a person approved, and the decision itself, both leave a trace. Changes go through Auto Mode and approval, and approval decisions land in the log.
Records by name, never by value. A log that stores secrets becomes a target. Credential events carry the key name only.
Honest limits
Retention is 90 days by default. Older entries are deleted automatically. If your policy needs a longer record, export to CSV before the window closes.
Plan-gated. Audit Logs are on Scale, Scale+ and Enterprise. Other plans see an upgrade prompt instead of the timeline.
Owner and Admin only. A workspace admin who isn't an organization Owner or Admin can't open it.
The guarantee is about the app. Entries can't be edited or deleted from the app, and we don't claim more than that here. Keep your cloud provider's own audit trail running alongside. The two answer different questions.
Try it
Give an agent one small, reversible write in a staging workspace, approve it if it pauses, then go find the receipt.
"Restart the checkout deployment in the staging cluster and tell me what you ran."
Open Admin Settings → Audit Logs, set the actor type to Agent and the date range to today, and read the Agent Actions tile. If you approved it, switch to the Approval category and find your own decision.
Start a trial or book a walkthrough.