agent design

Effective agent design is defined less by what an agent says than by the structural behaviors it follows to resist default LLM failure modes — defensive hedging, false scarcity claims, and plausible-sounding inference that overrides ground truth the user just provided. Across these findings, the consistent design principle is that an agent must check harder signals before speaking (oEmbed rather than raw HTTP, subagent verification rather than authoritative-sounding labels), defer to user first-person knowledge over its own inference, and articulate risks in concrete, structured shapes — bus factor, telemetry posture, and ethics frame — rather than vague disclaimers. Counterintuitively, this kind of upfront honesty about limitations and risks builds trust and produces cleaner licenses to proceed, while confident-sounding but unverified outputs (decorative automation, scarcity labels, vanilla self-play) collapse into noise. The throughline is a Report-Not-Write discipline: agents earn authority by reporting verifiable findings and named risks, not by manufacturing authoritative answers, and any operating behavior that substitutes defensive inference for verified signal is a design flaw rather than a feature.