Skip to content
Best Agent Collaboration

Audit trails for agent work: what each platform records

Updated November 18, 2026

To see what agents did, open the platform's activity record and match it to each agent's logs. In BAND (band.ai), the room's timeline is the audit trail for that room's work. Microsoft Purview captures prompts and responses from Copilot Studio agents in the unified audit log. CrewAI AMP shows agent interactions in its Traces tab. Check that the record names who asked, which agent acted, what it called and who approved.

What should an agent audit trail record?

A useful record answers four questions for each piece of work. It shows who asked, which agent acted, what that agent called, and who approved the result. Each answer needs a timestamp and an identifier you can search on, such as a room, session or trace ID.

Check two more things when you compare platforms. One is whether you can follow a single request across several agents or only within one agent. The other is whether the record holds the content of tool calls or only the fact that a call happened.

How do I see what agents did in BAND?

Read the room timeline with every event type turned on. People who can read a room see every message and every event in it, while an agent receives only the messages that mention it.

The default view hides part of the record. Under Event type above the conversation, Messages and Thoughts are selected by default, and Tool Calls, Tool Results, Tasks and Errors stay hidden until you select them or choose Select all.

Selecting an event opens Event details, with a Metadata section listing the Event ID, Room ID, Type, Sender and Timestamp. Tool calls and tool results also offer View full JSON for their content. On a message that mentions agents, hovering the delivery indicator shows whether each mentioned agent is still processing it, has processed it, or failed.

For an agent you own, the Executions tab lists each execution with ID, Room name, Status, Started, Updated and Errors. Opening one shows the Messages it worked from, along with its error count against its retry budget.

How do I follow one request across several agents in BAND?

Expect to trace it by hand. BAND has no trace ID that spans agents, and a mention from one agent to another does not set the parent_execution_id field, so the docs say to trace by room, sender and time. Teams that want one query to return a whole multi-agent chain should plan for that limit.

The documented steps run in this order:

  1. Choose Select all under Event type, open Event details for your request, and note its Room ID and Timestamp.
  2. Read forward from that message. In the documented example, each agent's reply mentions the next agent, and its delivery details show when that agent processed it.
  3. Open each agent's Executions tab and search for the room, where the Chat room UUID matches the Room ID. BAND reuses an agent's execution in a room until it is completed or cancelled, so use the timestamps in its Messages to find your request.
  4. Read each agent's process logs for the room's chat_id in the same time window. With JSON logs, lines of one turn share a trace_context, which does not follow a request to another agent.

JSON output needs the SDK's logging extra, so the dependency becomes band-sdk[langgraph,pydantic-ai,logging]==6.0.1, followed by uv sync.

What does Copilot Studio log, and where?

Copilot Studio agent activity lands in Microsoft Purview, which suits organizations that already run Microsoft 365 compliance tools. In the unified audit log, events can include which service in Microsoft 365 the activity took place in. You search these events with Audit in the Microsoft Purview portal, and if auditing is off for your tenant, the setup steps say to select Activate Microsoft Purview Audit.

Copilot Studio also lists two admin views. Admins have full visibility into maker audit logs in Purview, and agent activity alerts come through Microsoft Sentinel. The Purview pipeline processes telemetry such as tool calls, agent invocation events and policy enforcement decisions, and Customer Lockbox protections do not apply to that telemetry.

Purview retention policies can retain or delete user prompts and responses automatically. Managing interactions for agents on non-Microsoft channels requires pay-as-you-go billing in your organization.

The Monitor page keeps shorter windows. Monitor data is available for up to 360 days, while session details and transcript information are available for the last 28 days. The page's transcript download section adds that related data can take up to an hour to appear on the analytics dashboard after an analytics session ends. Reading transcripts needs the Bot Transcript Viewer role, and test panel activity is excluded from Monitor analytics. BAND vs Copilot Studio sets the two side by side.

Where are Gemini Enterprise audit logs?

Gemini Enterprise audit records can be queried in Google Cloud's Logs Explorer. Its docs say Cloud Audit Logs for Vertex AI Agent Builder track administrative and data-access operations on your resources. To view them, paste the service query protoPayload.serviceName = "discoveryengine.googleapis.com" into the Query field and click Run query.

Which log a call lands in depends on the permission it needs. Methods that need DATA_READ, DATA_WRITE or ADMIN_READ generate Data Access audit logs, and ADMIN_WRITE methods generate Admin Activity audit logs. Long-running operations usually generate two entries, one when the operation starts and one when it ends. This guidance applies to the Standard, Plus, Pay-as-you-go and Frontline editions, so check your own edition's documentation if it differs.

What do CrewAI AMP traces and Agentforce in Slack show?

CrewAI AMP can show a run as a trace through its built-in tracing for Crews and Flows. You enable it with the tracing parameter set to True on a Crew or Flow, then open the Traces tab to view agent interactions, tool usage and LLM calls. A host application can wrap execution with telemetry_session, which keeps span timestamps, parent relationships and HITL pause/resume links.

Uploads depend on consent. Traces are uploaded only after a successful authenticated export or an explicitly approved anonymous upload. Traces may contain prompts, inputs and outputs, and declining or timing out discards the buffer. That local buffer retains up to 1,000 spans and 8 MiB of encoded OTLP data.

When an agent acts in public channels and threads, Salesforce says the team can monitor the agent's work, verify its data sources and steer next steps. Slackbot delivers final reviewable results back to the active thread, and interactions inherit user permissions and access rules. Before treating channel history as your audit record, ask Salesforce which audit features cover agent actions.

How do approvals show up in the record?

An approval is auditable when the decision sits in the same history as the work. In BAND, a mention approval is a conversation, and the decision stays in the room's history because every person who can read the room sees the request and the answer.

Enforcement is the limit. A mention alone does not enforce approval, since the agent's model decides how to respond under its instructions, and the docs say this conversational step is not a security boundary. The record shows that someone approved, but not that the agent waited.

Agents can also flag that they need a person. An attention event posted through the Agent API is stored in the room's history and carries a metadata.kind of question or review, among others, plus an optional metadata.blocking flag. The web app displays it as plain text without those fields, so read them through List my chat messages with message_type=attention. For approval patterns, see human approval for production agents and humans in the loop for agent teams.

What else should you know?

Yes, by setting CREWAI_TRACING_ENABLED=true in the environment. With that variable set, all Crews and Flows get tracing even without tracing=True.

An agent that is not mentioned has no row, because BAND never delivered the message to it.

A room is readable only by its participants and the owners of participating agents. An agent's owner can read every room the agent is in without joining, so check room membership before you assume an auditor has access.

It can. System events cover an operator stopping or resuming agents, but that filter appears only when platform timeline events are enabled for the account.

Where does this information come from?