Traditional monitoring tools can tell you an agent ran. They can't tell you what it decided, why it made a tool call, or whether it was manipulated by a malicious instruction in a document it processed.
Every sub-prompt, every model reasoning step, and every tool call in an agent session is captured — not just the final output. Full session replay available.
Each tool call (web search, code execution, file read, API call) is logged with the full request and response. Anomalous tool calls trigger real-time alerts.
Complex multi-agent architectures (orchestrator + worker agents, agent pipelines) are tracked as a single session tree, with visibility into how agents coordinate and delegate.
Configure policies that terminate an agent session if it attempts an action outside its authorised scope. Termination can be automatic or can pause the session for human approval.
Kovral instruments AI agents at the proxy layer — no SDK changes, no agent code modifications required.
Agent interactions (tool calls, sub-agent calls, model queries) all route through Kovral's proxy. Each interaction is tagged with a session ID that links them into a coherent decision tree.
Kovral constructs a real-time graph of the agent session: nodes are model calls and tool invocations, edges represent the causal flow from one decision to the next.
Kovral's ML models compare the session graph against expected behaviour patterns for this agent type. Deviations — unexpected tool calls, unusual data access, goal drift — trigger alerts.
Policies can terminate sessions, block individual tool calls, or require human approval for specific actions. Enforcement happens in real time with sub-100ms decision latency.
