Abstract

A coding agent’s harness, the software that mediates the agent’s interaction with the world, determines how much of a model’s capability translates into delivered engineering work. As models have grown more capable, Claude Code’s harness has evolved alongside them. This talk is a retrospective on Claude Code’s public releases since the start of the year, detailing the design considerations behind each and the aggregate patterns they reveal. We frame each agent session as a trajectory: a search through the user’s environment that accumulates knowledge and enacts changes within it. Through that lens, the year’s launches sort into two moves: letting trajectories run further per unit of human attention (auto mode, routines, remote control, worktrees), and managing what each trajectory is conditioned on (memory, multi-agent code review). Meanwhile, increased model capability strengthens the trajectories themselves. We conclude by examining how to orient these trajectories toward desired outcomes. Once step-by-step correctness is no longer the constraint, the binding problems become reconciliation—merging many independently correct trajectories into one repository—and validation—making it cheap for humans to review and revise intent for agents that are competent but not clairvoyant. Both can be software engineering problems, and we offer them to this audience as open questions.

Speaker Bio

Dickson Tsai is a Member of Technical Staff at Anthropic, where he has worked on Claude Code since shortly after its launch. He pioneered hooks and built the foundations for skills support—Claude Code’s core extensibility primitives. Before Claude Code, he worked on pre-training data research at Anthropic and spent six years as a growth engineer on Google Search. He is currently interested in the right extension points for developer teams to run richer coding agent trajectories.