Agentic AI: Lessons from Deploy 2026 Panel
Implement measurement infrastructure and human verification loops for agentic AI to prevent costly errors and ensure reliability.
Set up a monitoring system that logs every agent interaction and schedule regular human reviews to validate outputs.
Summary
At DigitalOcean’s Deploy 2026 in San Francisco, a panel moderated by Dinesh Murthy explored what truly differentiates a production‑ready agent from a demo. The discussion featured Angela Hoover of Andi AI, Alex Mashrabov of Higgsfield AI, Hovsep Seraydarian of LawVo, and Peter Elias of Probably, each sharing how they built agentic systems that scale. Key takeaways included the necessity of creative DNA—bringing non‑technical creatives into the engineering loop—, the importance of human‑in‑the‑loop verification in high‑stakes domains like law, and the need for robust measurement infrastructure that logs every agent action for continuous self‑evaluation.
The panel also highlighted that “agentic” does not equal autonomous; LLMs must be prompted and cannot self‑check, so human oversight remains essential. Model selection is framed as a four‑variable trade‑off among cost, latency, intelligence, and capacity, with users demanding lower latency than founders expect. Finally, the founders warned that regulatory constraints and the risk of costly mistakes make it imperative to build guardrails and iterative testing into agentic products.
Key changes
- Creative DNA: involve non‑technical creatives in development
- Human‑in‑the‑loop verification is essential for high‑stakes domains
- Robust measurement infrastructure logs every agent action for self‑evaluation
- Agentic does not equal autonomous; LLMs require prompting
- Model selection is a trade‑off among cost, latency, intelligence, and capacity
- Users are more latency‑sensitive than founders expect