Lore

The Agent Line

Overview

Ankur Goyal (CEO, Braintrust) frames the boundary between work that should stay with a human and work that should be handed to an AI agent as "the agent line" — a threshold that keeps rising as models and agent scaffolding improve. Anything below the line is a candidate for delegation; because the line keeps moving up, people should periodically re-audit which of their current tasks have quietly fallen below it.

Mechanism

Rather than asking "can AI do this at all," the test is whether an activity would produce the same outcome if handed to an agent instead of done manually or discussed in person. Goyal's own example: benchmarking and algorithm exploration for a hard engineering problem (see Agent-Driven Exhaustive Experimentation (Practical vs. Theoretical Quality)) used to require a human staff engineer's sustained attention; today it sits below the agent line because an agent can run more rigorous, exhaustive experimentation than a person manually would.

Application

Audit routine decisions and interactions and ask which ones would produce an equivalent outcome if delegated to an agent rather than done by a person; move those below the line and reclaim the freed attention for work that still requires human judgment. See Evals as the Modern PRD ("What Not How") for how to hand off that delegated work well — via quantified success criteria rather than step-by-step instruction.

Linear: Shrinking Human-Owned Edges

Linear's team treats the boundary of what a human must still own in agent-assisted work as something that will keep shrinking as models improve, and builds roadmap around that assumption rather than around the current boundary: near-term work on proactivity and longer-running memory is explicitly aimed at letting the agent progressively own more of a project's full lifecycle, not just isolated tasks inside it. This is a working assumption more than a proven claim — the team treats it as the premise the roadmap is designed around, not a result they've already demonstrated.