Attention interface
What it is
An attention interface is the part of an agent system whose job is no longer running tools or managing memory, but explicitly managing the human's attention: deciding when a running agent may interrupt you, when it should keep working unattended, and which decisions it must route back to you and when. It treats a person's synchronous attention as the scarce resource the system has to budget, rather than compute or context length.
Why it's emerging now
As models absorb more of what used to be bespoke harness work — tool selection, memory, planning — directly into their own weights and policies, that work stops being the differentiator for the surrounding system. At the same time, agents increasingly run long, unattended background tasks while the human who launched them is doing something else. That combination leaves a genuine open problem: how does a background agent get a person's attention at the right moment, and how does a person reach back into an agent's ongoing work without babysitting it? Two independent groups converged on naming this within about a week of each other in August 2026 — one as an academic benchmark for a bidirectional human-agent coordination layer, the other as a practitioner's argument for where harness design is heading next.
How it relates to what you know
It's a successor framing to "Agent harnesses" and "Harness engineering" (Agent = Model + Harness) already on the ledger: as models absorb the classic responsibilities of a harness, this names what's left over once they do — a layer mediating between a long-running autonomous agent and a human's limited, intermittent attention. It also reframes "Human-in-the-loop" from a fallback safety checkpoint into an explicit, first-class design surface: not just "can a human approve this," but "when and how does the system decide it needs a human at all."