2026-04-05 · 14 pages
Most agent builder interfaces optimize for control: nodes, schemas, JSON tool definitions, and verbose system prompts. Control is the wrong target. The right target is time-to-first-running-agent, and the steepest variable in that equation is keyboard friction.
This paper documents the design principles behind GO Pilot GO's voice-first builder.
Principle 1: One utterance, one agent. The user should be able to describe the agent in a single breath. If they need a second utterance, it should clarify, not extend. We measure utterance length in seconds, not tokens; the median agent in our pilot data was created from a 9-second utterance.
Principle 2: Defaults beat fields. Every field that a UI exposes is a decision the user has to make. Voice-first design pushes those decisions to opinionated defaults (timezone, schedule cadence, retry policy, memory window) and only surfaces them when the user's utterance contradicts the default.
Principle 3: The agent earns its complexity. New agents launch with no memory, one tool, and a tight loop. Memory, additional tools, and recursive sub-agent calls unlock as the agent demonstrates usage, not at creation.
Principle 4: Voice is for description, text is for correction. Once the agent exists, edits should be available in both modalities. Operators report that re-describing an agent by voice is faster than editing, but text-based correction is essential for typo-level fixes and copy-paste of identifiers.
Principle 5: The transcript is the spec. Whatever the user said becomes the durable, human-readable description of the agent. The compiled system prompt is an implementation detail, not a deliverable.
Failure modes we have observed: voice transcription errors on brand names (mitigated by a per-workspace pronunciation dictionary), ambiguity in scheduling language (mitigated by an explicit confirmation step), and over-scoping in a single utterance (mitigated by a soft cap that prompts the user to split into two agents).
The full failure-mode catalog and our handling for each is included in the appendix.