A conference talk by Tejas Kumar (IBM) explaining what AI agent harnesses are and why they matter. An agent harness is everything around the model that grounds it in a stable, deterministic environment — including a tool registry, context management primitives, guardrails (e.g., max steps), an agent loop, and a verify step. The talk includes a live demo building a minimal browser-use agent that upvotes a Hacker News post using GPT-3.5 Turbo, incrementally adding harness components: guardrails to cap iterations and compress context, a verify step to detect lies/failures, and a deterministic login handler that injects credentials when the agent hits a login page. The key insight is that improving agent reliability doesn't require better prompts — a well-built harness can make even a weak model succeed at complex tasks.
Sort: