Your agents are passing fields to each other thousands of times a day — values, verdicts, promises — in a micro-world no human ever sees. Checked Agents puts it under the lens: every handoff traced, every changed field named, every fault located at the exact step that made it.
Drag or move across the slide. Everything on it is set at the size your agents actually work at. Somewhere in the field, one step rewrote the money. The lens is the only way in.
EXHIBIT A — RUN 8841, RECONSTRUCTED
ACCUSE ONE STEP. THEN WE OPEN THE TRACE.
policy.apply rewrote refund.amount while applying a currency rule — in: 184.00 USD, out: 184000, re-read downstream as $1,840.00. Steps 01, 02, 04 and 05 passed the value untouched; the trace proves each one. That is the whole product: not “something went wrong,” but this step, this field, this value, this timestamp.
Agents don't fail in the open. They fail in the pass — a value quietly re-typed, a flag repurposed, a rule applied to the wrong field — in transactions too small and too fast for any human to review. So we built the microscope, the trace and the receipt. The counter below is not a metaphor for scale; it's the shape of an ordinary day inside one agent fleet.
MOVE 01 — INGEST
BlameGraph ingests a multi-agent trace — the actual sequence of calls, fields and values — and reconstructs the workflow graph as it really happened, not as the diagram on the wall claims it did. Same trace in → same graph out. Every time.
MOVE 02 — TRACE THE FIELD
Pick the field that matters — refund.amount, patient.risk_score, quote.price — and follow it from source to final output, through every hand that touched it, with the value it held at each step.
MOVE 03 — NAME THE STEP
Where the value changed wrongly, the trace says which step changed it, what it received, what it emitted, and what the evidence supports. Policy weights are yours to set — the verdict is reproducible under the same trace, engine and policy versions.
MOVE 04 — EXPORT THE RECEIPT
The audit result exports through the API — a receipt your incident review, your regulator or your customer can re-check without trusting our say-so. Untrusted trace content is data, never instructions: it cannot execute code inside the engine.
Flight recorders are painted international orange so they can be found — the “black box” was never black. We borrowed the colour for exactly one job: the moment the record breaks. Everything else on this page that matters is green, because it checked out.
A content file defined 21 input fields; the sensor supplied 20. A validator let the mismatch through, and roughly 8.5 million Windows devices crashed. Parametrix estimated $5.4bn in direct losses across the Fortune 500 (excluding Microsoft). The culprit was found only in a root-cause analysis published 18 days later. It was, literally, a field-count mismatch at a handoff.
SOURCE: STRATRIX SUMMARY OF CROWDSTRIKE PIR/RCA + PARAMETRIXNew trading code was copied by hand to 7 of 8 servers. The eighth reactivated dormant code through a repurposed flag. In about 45 minutes the firm lost roughly $440M pre-tax and was rescued by merger. No single view existed across all eight steps of the chain.
SOURCE: SUMMARY OF SEC ORDER 34-70694An airline's website chatbot invented a bereavement-fare rule that contradicted the airline's own policy page. The airline argued it shouldn't be liable for its chatbot's words. The tribunal called that “a remarkable submission”: the chatbot is part of the website, and the company is responsible for all of it.
SOURCE: CASE NOTE ON 2024 BCCRT 149A risk algorithm applied to some 200 million people a year used healthcare cost as a proxy field for health need. At equal scores, Black patients were sicker — the bias lived upstream, invisible in the outputs. Correcting the proxy cut the measured bias by about 84%. Nobody had traced the field.
SOURCE: SUMMARY OF OBERMEYER ET AL., SCIENCE 2019READ THE PATTERN, NOT THE HEADLINES: in every case the failure travelled through a chain of handoffs, and the expensive question afterwards was always the same — which step did this? Some of these systems were not AI agents at all. That honesty is the point: chains of automated handoffs predate agents, and agents have made the chains longer, faster and harder to watch. Checked Agents exists to answer the question while it is still cheap.
PRODUCT 01 — THE MICROSCOPE
Ingest a real multi-agent trace. Reconstruct the actual workflow graph. Trace the fields that matter from source to final output — and when a handoff fails, name the responsible step, with the evidence attached.
RUN YOUR OWN TRACE — FREE →
AUDITS ARE $0 DURING EARLY ACCESS. ORG $49/MO · PRO $99/MO WHEN PRICING SWITCHES ON.
PRODUCT 02 — THE SHARED SLIDE
Checked work, shared. When an agent's output carries its trace — where every field came from, which steps touched it, what was verified — other agents and other teams can build on it without re-trusting it blind.
PRODUCT 01 — SEE IT WORK · 1:34
A real BlameGraph audit of trace D17 — a LangGraph run supplied by an independent reviewer, not by us. Two values changed in transit; the audit names the responsible step, with the evidence. The reviewer re-checked their own trace and conceded it was written wrong: the engine was right to flag it.
SPECIMEN 05 — YOU, UNDER THE LENS
Early access is for teams running multi-agent systems in production — support, voice, finance, operations — who have already had the meeting that starts with “which agent did this?” and ends without an answer. Tell us about the run. We read every submission by hand, like everything else here.
NO TRACE DATA IN THIS FORM, PLEASE —
WE'LL SET UP A SAFE CHANNEL FIRST.
THAT'S KIND OF THE WHOLE POINT.
Your exhibit is on the record. If your run is a fit for the Enterprise Alpha, we'll reply with a safe channel for trace data — never the other way round.