Agent UX Steering Affordances
UI controls that expose tool activity, let users edit or authorize calls, and regenerate from a corrected point so humans steer the agent mid-trajectory.
Conversational agents are assisted agency: the human stays in the loop to correct drift. Beyond a chat transcript, the harness should expose:
- Processing and tool-use indicators — show that work is in flight and that background tools ran.
- Tool-call detail — name, arguments, and results so users can inspect why a reply happened.
- Editable tool arguments — let users correct a bad call and resubmit.
- Regeneration from a corrected point — after a corrected tool call or message, regenerate the trajectory forward so steering sticks.
- Authorization prompts for any tool that might modify real-world assets — application-layer interception, not prompt-level pleading (human approval gates).
- Artifact visibility — show which conversation artifacts the agent currently sees, and allow dismiss when focus is wrong.
These affordances are harness design for LLM agents: they make tool traces and context selections inspectable and reversible. Chat turn-taking alone does not give users enough leverage when tool side effects or wrong attachments derail the task. In multi-agent simulacra (generative agent architecture), also offer an inner-voice channel — the user speaks as the agent’s conscience so directives stick more reliably than as an outside character — and let users rewrite nearby object state in natural language for the agent environment tree to notice.
Sources
- Prompt Engineering for LLMs (Berryman), ch. 8