Descriptions:
Arjun Singh, co-founder of Superconductor and previously of GradeScope (acquired after reaching millions of university users), shares six lessons from his team’s hands-on experience deeply integrating coding agents into a real product engineering workflow over the past year. The talk stands out from typical agentic tooling content by centering the humans collaborating around agents rather than the agents themselves.
Key lessons include: be model- and harness-agnostic, since the best model changes weekly and open-weight models like GLM 5.2 have proven competitive at lower cost; turn every human interface into an agent-and-human interface so that a single agent session persists across Slack, desktop apps, mobile, and GitHub with shared context rather than being siloed to one tool; and make agent work visible and collaborative so the whole team can observe, comment on, and redirect ongoing agent sessions in real time. Singh also covers eliminating “lid anxiety” by running agents in isolated cloud environments so work continues when laptops close — a pattern his team adopted after he found himself tethered to his laptop with a six-month-old at home.
The talk closes with a lesson on using customer calls and team meetings as a live idea feed: conversation transcripts flow into the agent system, which automatically creates and begins prototyping tickets, often producing shippable PRs with minimal human intervention. Superconductor’s product enables several of these patterns, but the lessons are framed as broadly applicable principles for any engineering team building collaborative workflows around agentic coding tools.
📺 Source: AI Engineer · Published August 09, 2026
🏷️ Format: Keynote Launch
