Skip to main content
The Oso for Agents UI is a control plane for understanding what your agents actually do in production and tightening permissions and controls based on observed behavior. Use it to:
  • See agent activity end-to-end (actions, decisions, context)
  • Detect misbehavior early (before it becomes an incident)
  • Enforce guardrails at the action boundary (approvals, rate limits, information-flow controls)

​
Detecting Misbehaviour

​
Types

The UI is optimized to surface issues that matter for agent workloads:
  • Velocity spikes: Agents suddenly executing far more actions than normal (often a sign of a loop, prompt injection, or mis-scoped automation).
  • Data leaks: Risky sequences like accessing sensitive data and then attempting an outbound or state-changing action.
  • Intent drift: The agent is “still working” but is now using different tools, hitting different resources, or acting outside the expected workflow shape.
Recommended use:
  • Start from a flagged agent or alert.
  • Drill into the recent action timeline to see the exact tool calls and parameters.
  • Compare it to historical behavior to identify the root cause of the misbehaviour.

​
Simulations (Pre-Release Testing)

Before enabling a new tool or workflow in production, use simulations to validate:
  • Which tools the agent attempts to use
  • What permissions are actually exercised
  • Whether the workflow triggers high-risk sequences

​
Production Monitoring

In production, the UI becomes your operational loop:
  • Track what changed (new tools, new resources, new patterns)
  • Identify over-privilege (permissions that are present but not used)
  • Triage high-signal anomalies quickly with action-level evidence (not prompt transcripts alone)

​
Enforcing Controls

​
Approvals / HITL

When a tool call is risky, enforce human approval before execution. In the UI, reviewers should be able to see:
  • The tool being invoked and parameters
  • Why it was flagged (policy or risk signal)
  • Relevant prior steps in the session (the minimal context needed to decide)
Recommended use:
  • Start with approvals on high-impact tools (writes, outbound comms, admin actions).
  • Expand coverage based on what you observe.

​
Rate Limiting

Rate limits mitigate blast radius when agents misbehave:
  • Cap action volume per agent/session/tool
  • Block or degrade gracefully during spikes
  • Create predictable failure modes instead of runaway execution

​
Information Flow

Information-flow controls help prevent unsafe combinations across a session, for example:
  • Untrusted input → sensitive data access → outbound or state-changing action
Use the UI to:
  • Understand which step introduced risk
  • Confirm the sequence that triggered enforcement
  • Tune policy so safe workflows pass while risky compositions get gated

​
Incident Response & Quarantine

When something goes wrong, the UI should support a clean incident loop:
  • Detect: alerts triggered by executed actions and abnormal patterns
  • Investigate: action timeline + tool call inspection (what happened, where, and with what parameters)
  • Contain: quarantine an agent, require approvals, tighten limits
  • Learn: turn the incident into a durable policy or control (so it doesn’t recur)