The Eight Frame Agents

Eight specialists. One of them exists to prove the others wrong.

Trace8 doesn’t run one model with a system prompt and a confidence score. It runs a pipeline of eight specialized agents — each with a defined job, a structured contract, and a single responsibility in the chain from alert to decision.

01 — TRIAGE

The opening verdict

First contact with the raw alert. Normalizes, assesses, and returns a structured call — true positive, false positive, or escalate — with severity and reasoning attached. Everything downstream builds on this contract.

Reads: raw alert  ·  Outputs: structured verdict + severity
02 — HUNT

The evidence gatherer

Pivots across your log sources for everything relevant — the facts that support the verdict and the facts that undercut it. Hunt is deliberately not partisan. It builds the full picture so Challenge has something real to attack.

Reads: log sources  ·  Outputs: corroborating + contradicting evidence
03 — CHALLENGE

The adversary

The agent that makes Trace8 trustworthy. Challenge takes the Triage verdict and tries to break it — searching for the benign explanation, the missed context, the reason the confident answer is wrong. Runs on every verdict, without exception. Adversarial validation is the architecture, not an add-on.

Reads: verdict + evidence  ·  Outputs: validation or dissent
04 — EVIDENCE

The case file

Assembles the investigation into something auditable: what’s known, what’s inferred, what’s still missing. The gaps are first-class output — an honest investigation names what it couldn’t confirm.

Reads: full investigation state  ·  Outputs: structured case file
05 — EDR MANAGER

Host-level ground truth

Reaches into endpoint telemetry to confirm or contradict the log-level story. Process lineage, parent-child chains, signing status — the detail that separates a real compromise from a look-alike.

Reads: endpoint telemetry  ·  Outputs: host-level confirmation
06 — EMAIL GUARDIAN

The delivery vector

Owns anything email-borne. Resolves sender, payload, and the full delivery path to answer the question every phishing investigation turns on: how did this get in, and what did it carry.

Reads: email telemetry  ·  Outputs: vector resolution
07 — BRIEFER

The human handoff

Translates the machine investigation into a human decision. Not a field dump — a narrative: what happened, why the system believes it, and what to do next.

Reads: completed investigation  ·  Outputs: analyst-ready narrative
08 — AUDITOR

The final gate

Validates the whole chain before anything surfaces. Internal contradictions, broken reasoning, incomplete evidence — caught here, before they reach a human. The last check between the pipeline and your screen.

Reads: full agent chain  ·  Outputs: pass / flag

Why eight, not one

The architecture is the product.

A single large model can do all of this in one pass. It just can’t be audited when it does. Separating the work into eight agents with structured contracts means every step is inspectable, every handoff is clean, and the adversarial check is a real, independent process — not the same model grading its own homework.

Watch all eight run on a real alert.