Skip to content

Captured Events

Every action AI Agent Watch observes on a host is reported as an event. Events belong to a session (an agent_session_started/agent_session_end pair, identified by a session ID) and share a common set of fields, plus fields specific to that event type.

Common fields

Every event carries, among others:

  • App, host, agent type/version/process, and the session ID it belongs to
  • Event type and a human-readable message
  • Is trusted agent - whether the agent that produced the event is trusted
  • Priority and risk score - set by whichever alert rule matched the event, if any
  • Matched rules - which rule(s) matched, if any

Event types

Event type What it represents
agent_session_started An AI agent process/session began. Carries the command line used to start it.
agent_session_end The session ended. Carries the session's rolled-up totals - duration, session risk score, number of tool executions, connections, requests, file reads/writes, and bytes sent/received.
tool_executed The agent spawned a tool, shell command, or subprocess. Carries the command, its full command line, and the spawned process' PID.
tool_exit A previously spawned tool/subprocess exited. Carries its exit code.
file_read The agent read a file. Carries the file path and extension.
file_write The agent wrote to a file. Carries the file path and extension.
file_rename The agent renamed or moved a file. Carries the old and new paths.
net_connect_started The agent opened an outbound connection. Carries the destination host/IP/port and whether that destination is a trusted host.
network_activity A rollup of traffic to a given destination. Carries bytes sent/received and request count.
pii_detection Outbound data matched one of your PII categories. Carries the category, the masked/one-way-encrypted payload (see PII Categories), the HTTP method/URL, and the destination.
exec_failed The kernel blocked the agent from executing a binary - due to an existing OS-level permission or security policy, not anything AI Agent Watch enforces. Carries an error code/reason.
file_open_failed The kernel blocked the agent from opening a file, for the same reason. Carries the file path/extension and an error code/reason.
exec_denied A governance policy blocked tool execution based on a binary or command-argument match.
file_access_denied A governance policy denied file access based on a path or extension.
net_denied A governance policy blocked network activity based on an IP, port, CIDR, or destination-host match.

exec_failed and file_open_failed report denials caused by existing OS permissions or security policies. exec_denied, file_access_denied, and net_denied report AI Agent Watch Governance enforcement. See Governance for policy configuration and notifications.

Governance events record the matched condition for the blocked operation. Depending on the event type, they also carry the binary, file path, or destination IP and port.