Harness → attention interface: human attention as the scarce resource
Dan McAteer on Latent Space traces harness eras from AutoGPT to Claude Code, then argues what’s left after models absorb scaffolding is an interface to human attention.
Source: The Evolution of the Agent Harness
Narrative arc: harness asks vs model can deliver. AutoGPT overshot; IDEs pulled autonomy back; Claude Code crossed when models were ready. Harness-Bench’s spread on the same model shows half the agent is the harness. RL inside harnesses, then absorption (compaction trained in), then deletion of scaffolds.
Punchline for builders: keep deleting what models absorb. What’s left is permissions, identity, trust, legibility—the harness as the agent’s interface to scarce human attention. Predicts an “attention policy surface” analogous to AGENTS.md: when to interrupt, what can auto-decide, what needs approval.
Quotes the OpenAI harness thread: synchronous human attention is the scarce resource once tokens are abundant.
Why it matters for code review
Review queues are attention interfaces whether we admit it or not. Designing interrupt policy, escalation, and legible evidence is the next harness layer after “catch more bugs.”