Case 02 / Agent behaviour · Human control
Designing for agency
A trustworthy agent experience needs more than a reassuring interface. It needs clear decisions about initiative, uncertainty, permission and recovery.
- My contribution
- Initiated and authored behaviour standards, shared prompt guidance and evaluation scenarios.
- Scope
- Agent interactions, behaviour specification and evaluation design.
- Concrete outcome
- Design principles expressed as observable, testable expectations.
01 / The problem
What does “trustworthy” mean in an actual interaction?
As agents take a more active role in a product, the interaction depends on more than the screen. What does the agent claim? When should it ask? What happens when the person corrects it?
High-level principles leave room for inconsistent interpretations. I worked on turning those principles into specific behaviour that a team can implement and assess.
02 / The approach
Connect the principle, the instruction and the evaluation.
I developed guidance for explaining recommendations, communicating uncertainty, asking for permission and responding to corrections. Those decisions became behaviour standards and shared prompt guidance.
I also specified evaluation scenarios with expected behaviour, failure examples, deterministic checks and rubrics for model-based assessment. The aim was to give product, design and engineering a shared definition of what they were trying to build.
Decision 01
Tie statements to system state
An agent should only describe an action as complete when the system has confirmed it. When confirmation is missing, the response should preserve that uncertainty.
Decision 02
Treat correction as part of the task
A person should be able to correct the agent’s understanding. The response needs to acknowledge what changed and make the proposed next step clear.
Decision 03
Make standards maintainable
A reusable rule needs a concrete expectation and an evaluation scenario. A pattern observed once can remain a pattern note until there is enough reason to generalise it.
Explore the design principle
Situation: an action was requested, but the system has not returned a confirmation.
I haven’t received confirmation that the change was saved. I can check the current status before we continue.
Design principle: describe the confirmed state, and make the uncertainty and next step clear.
Illustrative dialogue written for this portfolio. It demonstrates the reasoning, not a production prompt, tested output or live AI system.
03 / The outcome
A concrete vocabulary for agent quality.
The work produced standards, prompt guidance and specified evaluation cases. It connected design intent to instructions and expectations that colleagues could discuss in concrete terms.
What I contributed
- Authored guidance for user-facing agent behaviour.
- Translated principles into shared prompt guidance.
- Specified expected behaviour and failure scenarios for evaluation.
What the evidence does not establish
These are concrete design deliverables. Current rollout, evaluation coverage and an operational evaluation harness are not verified here. No improvement in reliability or user trust is claimed as measured.
04 / The next question
What would I want to learn next?
The next evidence would combine observed agent behaviour with task outcomes and qualitative user feedback. A switch back to a familiar interface alone would not establish why someone lost confidence.
Continue exploring / Case 03
A fair reading of results