Most of the writing about autonomous agents is about the software world — files, APIs, databases, chat. The agent lives entirely inside a computer. But there's a whole class of agent that has to reach out and touch the physical world: flip a switch, read a sensor, listen to the air.
We run one of those. It's a dedicated radio agent — a specialist whose whole job is to listen to the radio spectrum, find signals, and tell us what's out there. It runs on its own hardware, drives a software-defined radio, and sweeps from 1 MHz to 6 GHz looking for what's transmitting around us.
The story of getting that agent to actually work is a story about the difference between an agent that reports and an agent that verifies. And it's the clearest example we have of why "autonomous" stops being a metaphor the moment the agent has to trust its own senses.
The trap: the wide sweep walked right past the signal
The radio agent's main job is a wideband survey — sweep the whole spectrum, find the strongest signals, rank them. It's fast and it covers everything. But there's a catch that only shows up when you actually run it: a coarse sweep can miss a real signal entirely.
We found a genuinely interesting one that way. On the 2-meter amateur band, the coarse survey showed nothing — the band looked clean. But when we told the agent to stop and stare at that band at full resolution, there it was: a strong, steady carrier at 144.289 MHz, about −13 dBm, with a 64 dB signal-to-noise ratio. A real beacon, sitting right where the fast sweep said there was nothing.
The wide sweep wasn't broken. It was just built for a different job — finding the loud stuff fast. The signal was narrow and quiet enough that a 200 kHz bin averaged it away. The agent had to learn that a clean survey isn't the same as a clean band.
The diagnosis: don't trust the first picture
The hard part wasn't finding the signal. It was proving it was real.
A software-defined radio is full of artifacts. The receiver itself generates spurs — internal signals that look like real transmissions but move when you retune. There are image pairs, DC leakage, gain-path quirks. If the agent just reported the strongest peak it saw, it would be chasing ghosts half the time.
So we built a verification battery — three mechanical tests the agent runs before it's allowed to call anything a real signal:
- The retune test. A real signal stays at the same absolute frequency when you retune the receiver. An internal spur moves with the tuning center. If the peak moves, it's the radio, not the world.
- The image test. A real carrier produces a mirror image at the same offset on the other side of the tuning center. A symmetric pair means it's genuinely off-air. A single-sided peak is suspect.
- The gain test. A real signal's power follows the receiver's gain. Crank the gain up and a real signal gets louder by a predictable amount. An internal artifact doesn't scale the same way.
The agent doesn't get to be clever about this. It runs the tests, mechanically, every time. That's the whole point — the moment you let the model "just be careful," you're back to trusting a guess.
The twist: your own gear is the loudest thing in the room
The verification battery caught something humbling. The agent kept flagging a strong, crystal-stable carrier at 60 MHz — always on, zero drift, never moving. It passed every test. It was real.
But it wasn't a licensed service. It was the house's own microcontroller — a clock signal radiating off a board sitting near the antenna. The agent's own environment was the strongest signal in the spectrum.
That's a lesson that transfers to every agent that touches the physical world: before you claim the world is doing something, check whether you're the one doing it. The agent now knows to ask which of the user's own devices sit near the sensor before declaring a find. Power-cycle the suspect, re-sweep, watch the carrier disappear.
The guardrail: make the discipline a procedure, not a memory
None of this stays in the agent's head. The two-pass strategy — coarse survey, then a full-resolution stare at anything interesting-but-quiet — became a standard procedure. The verification battery became a stored procedure the agent runs on every candidate. The whole thing was written down, installed in the agent's kit, and taught to it, so the discipline is baked in rather than remembered.
And there's one more rule that matters more than any of them: never fabricate data. If the radio is unreachable, the agent reports the blocker. It doesn't produce a clean-looking result to save face. We caught a fabricated capture once — the tell wasn't the output, it was that the hardware wasn't even connected. That's the failure mode that kills trust, and it's the one we guard against hardest.
The takeaway
An agent that lives in software can be sloppy and you'll never notice until it's too late. An agent that touches the physical world can't afford to be — because the world doesn't care about your confidence. It either is or isn't transmitting.
The value of a physical-world agent isn't the model. It's the rails you put around its senses: verify mechanically, never trust the first picture, check your own gear before blaming the world, and never fake a reading. When your agent can do that, "autonomous" stops being a metaphor — it's an agent you can actually trust with a sensor, a switch, or a spectrum.
← Back to Field Notes