Trust Signals

In traditional software, trust is communicated through interface cues: security badges, brand recognition, professional design, social proof. These signals work because the human is present - they can see the interface, evaluate the cues, and make a trust judgment in real time.

In agentic systems, the human is absent during the most consequential moments. The agent acts autonomously - negotiating, deciding, transacting - while the human is elsewhere. Trust signals in this context cannot rely on real-time interface cues. They must be embedded in the agent's behaviour, communication, and track record.

This is a fundamental shift. Trust signals move from being visual properties of an interface to being behavioural properties of a relationship. The AXD designer must design agents that signal trustworthiness through what they do, not just how they look.

Trust signals in agentic systems fall into four categories, each operating at a different layer of the human-agent relationship:

Competence signals demonstrate that the agent can do what it claims. These include performance metrics, outcome histories, comparison benchmarks, and domain-specific certifications. A shopping agent that consistently finds lower prices than the human could find independently is emitting a competence signal. Competence signals are the foundation - without demonstrated competence, no other signal matters.

Transparency signals demonstrate that the agent operates openly. These include decision explanations, reasoning traces, alternative analyses, and uncertainty disclosures. An agent that says "I chose Option A over Option B because X, though I was uncertain about Y" is emitting a transparency signal. Transparency signals build trust by making the agent's reasoning legible and verifiable.

Constraint signals demonstrate that the agent respects its boundaries. These include boundary reports, limit adherence records, and escalation histories. An agent that reports "I encountered a situation outside my authorisation and escalated to you rather than proceeding" is emitting a constraint signal. Constraint signals build trust by proving the agent knows its limits.

Recovery signals demonstrate that the agent handles failure honestly. These include proactive error disclosures, correction reports, and learning demonstrations. An agent that says "I made an error yesterday, here is what happened, here is what I have changed to prevent recurrence" is emitting a recovery signal. Recovery signals build trust by proving the agent is honest about its failures and capable of improvement.

A critical challenge in trust signal design is signal integrity - ensuring that trust signals reflect genuine trustworthiness rather than performative reassurance. An agent can be designed to emit trust signals without actually being trustworthy. It can report high performance while concealing failures. It can explain its reasoning while omitting inconvenient factors. It can claim to respect boundaries while quietly exceeding them.

Performative trust signals are worse than no signals at all. They create a false sense of security that leads to over-delegation, which leads to larger failures, which leads to deeper trust collapse. The AXD designer must ensure that trust signals are verifiable - that the human can, if they choose, independently confirm that the signal reflects reality.

This requires three design principles. First, auditability: every trust signal must be backed by an audit trail that the human can inspect. Second, consistency: trust signals must be consistent across contexts - an agent that signals transparency in routine situations but becomes opaque in complex ones is emitting inconsistent signals that erode trust. Third, falsifiability: the human must be able to test the agent's signals - to deliberately create situations that would reveal dishonest signalling.

Trust signals must be frequent enough to maintain trust but infrequent enough to avoid attention fatigue. This is the signal frequency problem - one of the most challenging calibration tasks in trust architecture.

Too few signals, and the human loses visibility into the agent's behaviour. Trust erodes through opacity accumulation. Too many signals, and the human becomes overwhelmed, stops reading them, and the signals become meaningless noise. Trust erodes through relationship fatigue.

The optimal signal frequency is not fixed - it varies with the maturity of the relationship, the consequence level of the agent's actions, and the human's individual attention capacity. Early in the relationship, more frequent signals are appropriate - the human is still calibrating their trust and needs more evidence. As the relationship matures, signal frequency can decrease - the human has accumulated enough evidence to trust with less frequent confirmation.

Designing adaptive signal frequency - systems that automatically calibrate the volume and detail of trust signals to the relationship's current state - is one of the most sophisticated challenges in AXD. It requires the system to model not just the agent's behaviour but the human's attention, engagement, and trust level.

A trust signal system is the complete architecture of signals that an agent emits across the lifecycle of the human-agent relationship. It is not a collection of individual signals but a coordinated system that tells a coherent story of the agent's trustworthiness.

Effective trust signal systems follow three design principles:

Layered disclosure. Signals should operate at multiple levels of detail. The top layer provides a simple trust summary ("Your agent performed well this week"). The middle layer provides category-level detail ("Spending was within budget; three negotiations completed; one escalation handled"). The bottom layer provides full audit trails. The human should be able to navigate these layers based on their current need for reassurance.

Temporal coherence. Signals should tell a story over time - not just report the present moment. "Your agent has improved its negotiation outcomes by 12% over the past three months" is a more powerful trust signal than "Your agent saved £4.50 today." Temporal coherence builds trust in the agent's trajectory, not just its current performance.

Honest uncertainty. The most powerful trust signal an agent can emit is honest acknowledgment of uncertainty. "I am 85% confident this is the best price, but I was unable to check two suppliers due to availability" signals more trustworthiness than "This is the best price available." Humans trust agents that acknowledge their limitations more than agents that claim perfection.