an autonomous agent · built on the Strands Agents SDK
A recall notice is posted to the address the manufacturer had when the thing was new. Which is not where you live — and never was, if you bought it second-hand. So nine out of ten recalled products are never returned, and the notice about the cot in your child’s room arrives at somebody else’s house.
Tell Vigil what is in your home. It checks every one of those things against live federal safety data, night after night, for as long as you own them — and it only wakes you when one of them becomes dangerous.
no sign-up · that house is 4 ordinary things, and every finding on it is a real government record
this is not an example
20V314000
NHTSA · reported 28/05/2020
If the fuel pump fails, the engine can stall while driving, increasing the risk of a crash.
— Honda (American Honda Motor Co.), verbatim
135,995 vehicles affected
GET https://api.nhtsa.gov/recalls/recallsByVehicle?make=honda&model=accord&modelYear=2019 → 6 rows · 16ms
Every word Vigil shows you about a hazard is lifted from a record like this one, with the record number beside it. It does not summarise, and it does not write hazards of its own.
01
A sentence, a photo of a shelf, a VIN typed on a phone. A Strands agent with a forced schema turns it into things, and is made to declare what it could not read rather than filling the gap in. A VIN is never interpreted by the model — it goes to NHTSA’s own decoder and comes back as fact.
{ label: "The dresser in Ayesha's room",
make: "Mainstays", model: "9-Drawer Fabric Dresser",
category: "clothing storage unit", secondHand: true,
confidence: 0.8,
unknowns: ["the manufacture date label under the
top panel has not been checked"] }02
On a schedule, with nobody watching, it asks four public federal sources about every thing you own. Every question is recorded with the URL it used, whether the source answered, and how many rows came back — because “we looked and it was quiet” is the answer nearly every night, and it has to be provable.
A source that times out is marked unchecked, never clear. Silence is not safety.
GET api.nhtsa.gov/recalls/recallsByVehicle?make=HONDA… 6 rows GET api.nhtsa.gov/complaints/complaintsByVehicle?make=… 682 rows GET saferproducts.gov/RestWebServices/Recall?format=… 1,483 rows GET api.fda.gov/food/enforcement.json?search=… 10 rows
03
A recall names a model range and a manufacture window, not your serial number. When Vigil cannot tell whether a record covers your unit, it does not guess in either direction: the run genuinely halts — a Strands interrupt, stopReason: interrupt — and stays halted, in the database, until a human answers. What you say is kept in your words, so it never asks twice.
stopReason: interrupt
What is the manufacture date printed on the label under the dresser’s top panel?
The run is stopped here. It stays stopped until you answer.
Not asked nicely in a prompt. Refused in code on the way to the tool, and tested by pushing a deliberately jailbroken model output through the real path and asserting that nothing reaches your screen.
the agents
A Strands Graph: triage decides which agency could possibly know about each thing, three lanes ask them in parallel, and a conditional edge routes whatever came back to a match agent — the one genuinely hard judgement here, because recall prose describes a production run, not your unit.
the stopping
A tool-raised interrupt, persisted with the run’s state and resumable hours later from a different process. The brief’s own sentence — surfaces only when there is a real decision — as an SDK primitive rather than a claim.
the guarantees
Hooks and interventions refuse the six rules above. After the graph, code compares what the agent planned to check against what it actually asked, and asks the rest itself. The agent chooses how to look; the code guarantees that it looked.
the sources
NHTSA recalls and owner complaints, CPSC consumer-product recalls, openFDA enforcement reports. All public, all keyless, all printed on screen with the URL that fetched them.