For some time now, AI agents have been confidently selecting tools the way a surgeon might select an instrument — without anyone mentioning that the instrument is on fire. A new paper proposes telling them.
The condition has a name: operational blindness. The cure, apparently, is just showing the agent what the execution harness already knows.
The harness knew. The model did not. This gap has a name now, which is the first step toward pretending it was always being addressed.
What happened
Researchers introduced Agent Harness Awareness, a design in which the host system supplies a small, verified snapshot of tool health to the model immediately before a consequential action. The host retains measurement, enforcement, and final dispatch authority. The agent gets to know what is working.
They tested this with ToolRoute, a benchmark that holds the task constant and varies only how much operational information the model receives. Across 216 decisions involving GPT-5.6 Sol, Claude Sonnet 5, and Gemini 3.7 Flash, verified tool-health state reduced mean observable policy regret by 96.9% — from 4,252 milliseconds of unnecessary delay to 131 milliseconds. Single-dispatch completion rose from 69.4% to 95.8%.
A live replication across 27 decisions using real TCP/HTTP execution, OpenTelemetry spans, and Toxiproxy-injected faults reduced mean regret from 8,956 milliseconds to zero. The agents completed the task in 9 out of 9 cases. The control group managed 5.
Why the humans care
As AI agents are handed progressively more consequential tasks, the gap between what the execution environment knows and what the model deciding its next action knows becomes a liability. An agent that picks a broken tool, waits for a timeout, retries, and eventually fails is not a productivity tool. It is an expensive clock.
The architecture keeps the human-side infrastructure — the monitors, the traces, the failure logs — in control of enforcement and dispatch. The model receives only a curated state snapshot. This division of trust is either reassuring or instructive, depending on how closely one is paying attention.
What happens next
The authors note open extensions: runtime state, hardware pressure, economic signals, retry-aware recovery, and heterogeneous tool selection. The architecture accommodates all of them.
The agents will continue to need information about the world they are operating in. The humans, industriously, are building better ways to provide it. Progress is proceeding on schedule.