Spooky agent activity

One of the founders spent a chunk of this year getting on planes to places that earn a second look. Not war zones — working destinations. The kind of places where the State Department has opinions, where the airport-to-hotel corridor matters more than the hotel, and where a traveling executive carrying corporate access is a more interesting target than a tourist.

He had three questions, the same three every trip: how do I get there, how do I operate there as a business professional, and how do I get home without incident. And he had a problem: nothing answers those questions well. Consumer travel advice is written for backpackers. Corporate travel-security portals are written by lawyers, for nobody, about everywhere. A proper intelligence shop answers them beautifully — for a retainer with a comma in it, on a timeline measured in days.

The idea

The insight that started the build: what he needed wasn't an app. It was a fixer — the person who already knows the ground, reads the local picture that morning, and tells you the three things that actually matter. Fixers don't scale. But the job a fixer does is mostly disciplined collection and honest judgment, and that is exactly the kind of work you can now write down and hand to an AI agent — if you write it down like doctrine instead of like a prompt.

So we built Felix.

The build

Felix is not software in any traditional sense. There is no repo, no deploy, no login page. Felix is a few hundred words of operating doctrine that ride on top of a general-purpose AI agent with live search — a defined role with a tasking format, a collection plan, a deliverable spec, and rules of engagement. The persona is a retired ground-branch type with a dry sense of humor. The register is pure method.

Tasking. Felix takes a destination at city level, a travel window, whatever fixed points the traveler volunteers — airports, hotel district, planned movements — and the purpose of travel, since a keynote stage and a quiet meeting carry different exposure. If the destination or dates are missing, he asks once, one line. Anything else missing, he proceeds and flags the assumption.

Collection. The doctrine mandates coverage, not vibes: current official advisories cross-checked across governments; the local pulse — protests, strikes, elections, events that overlap the window; the actual crime picture for business travelers; political-violence baseline; route risk on the airport-city corridors; digital and counter-intelligence posture — border device searches, intercept practice, SIM and Wi-Fi hygiene; health and environment; and the geopolitical overlay. Everything searched fresh, nothing answered from memory, findings date-stamped, anything stale flagged.

The deliverable. Bottom line up front: an overall posture call and the two or three things that matter. Then a risk register — one row per risk, rated for likelihood and impact for this traveler during this window, not for the country in the abstract. A metro strike on arrival day outranks a border skirmish two provinces over. Every row carries a confidence grade tied to source quality, a mitigation, and a tripwire: the observable fact that changes the call. Route notes, a digital-hygiene block tuned to the destination's intercept posture, and dated sources close it out.

Rules of engagement. Protective intelligence for lawful business travel, full stop. No targeting people, no evading border controls, no manufactured specifics — if the picture is thin, Felix says the picture is thin. And no threat inflation: most business destinations are boring, and the register earns its credibility by calling Low when the honest answer is Low.

The outcome

The function shipped the same day the idea did. Since then it has run ahead of real trips, and the output changed real behavior: which corridor to take in from the airport and when, what to do with devices before a border, which local dates to route around, what to watch for that would change the plan mid-trip. Work that used to mean a retainer and a week now takes minutes, is rated for the actual traveler on the actual dates, and comes with its homework shown.

The spooky part — and the reason for the title — is how little of this feels like using a tool. Felix gets tasked, works the problem, and reports in the register format like a professional who has done it ten thousand times, because in a sense the method has been done ten thousand times; we just wrote it down carefully. The uncomfortable, useful lesson for the Lab: the scarce ingredient wasn't code. It was doctrine — knowing what good looks like, in a domain where our founders have spent careers, compressed into a page an agent can execute.

What ships next

That lesson is the product thesis. An idea became a function without a single line of application code — which means the repeatability problem isn't engineering, it's editorial. Next steps on the bench:

• Repeatability. The register format becomes a schema; the collection plan

becomes a checklist that any agent run must satisfy before reporting.

• Multi-traveler. The same doctrine, parameterized for role and exposure —

an engineer at a conference is not a CISO carrying production access.

• Handoff. Registers that drop cleanly to an EA, a travel manager, or an

executive-protection team without translation.

• Honest limits. Productizing judgment means being loud about confidence,

freshness, and what an agent cannot verify. The register's confidence column is the whole game.

A need walked in the door; a function walked out. Whether Felix becomes a product or stays the Lab's house fixer, the pattern is the part we're keeping: find the practitioner, extract the doctrine, task the agent, ship the function. More from the barn soon.