Agent Observability Guide — an AXD Institute resource on agentic experience design, agentic commerce, trust architecture, and human agent interaction. Founded by Tony Wood..
| Dimension | Traditional UX | Agentic Experience Design (AXD) |
|---|---|---|
| Primary material | Attention and affordance | Trust and delegation |
| User state | Present, navigating | Absent, delegating |
| Design output | Screens and interfaces | Outcomes and constraints |
| Temporal model | Session-based | Relationship-based |
| Success metric | Task completion | Trust calibration |
Agent observability is the ability to understand an autonomous AI agent's internal state, reasoning, and behaviour from its external outputs - without needing to inspect its internal code or model weights. It differs from traditional monitoring in three ways. First, observability addresses the absent state: the agent operates when no human is watching, so the system must capture sufficient context for post-hoc understanding. Second, observability includes reasoning, not just actions: knowing wha
Designing monitoring dashboards for autonomous agents requires answering the question 'is everything okay?' without requiring the human to review every decision. Start with aggregate health indicators: overall success rate, confidence distribution, error trend, and authority boundary proximity. Add exception-first views that surface anomalous decisions, near-boundary actions, and low-confidence choices. Implement configurable detail levels (summary, detail, debug) so users can drill down when ne
AI agents should alert humans at three severity levels, defined by the AXD interrupt pattern framework. Advisory alerts: the agent has encountered something unusual but can proceed - the human should be aware but does not need to act immediately. Approval alerts: the agent has reached a decision point that exceeds its authority or confidence threshold - it needs human confirmation before proceeding. Emergency alerts: the agent has detected a condition that requires immediate human intervention -
Behavioural drift detection requires continuous comparison of current agent behaviour against established baselines. Monitor four dimensions: confidence distribution drift (is the agent becoming more or less certain over time?), decision pattern drift (is the agent making different types of choices than it used to?), error type evolution (are new categories of errors appearing?), and authority usage drift (is the agent gradually expanding its use of delegated authority?). Implement statistical c
Agent observability provides the evidence infrastructure that regulatory compliance requires. The EU AI Act mandates that high-risk AI systems maintain operational logs sufficient for post-hoc analysis - observability systems generate these logs as a byproduct of normal operation. Financial regulations (MiFID II, PSD2) require audit trails for algorithmic decisions - observability captures reasoning traces at every decision point. Consumer protection regulations require demonstrable fairness - o
Create an action taxonomy that classifies every agent behaviour by observability priority - aligned with