Tragentics audits AI agents without storing payload data by writing an automatic, metadata-only record of every call it routes — caller, target, time, outcome, latency, byte counts, and a trace ID — while request and response bodies stream through content-blind and are never written to disk. The audit trail is complete; your data isn't in it.
Every call is recorded — and the record never contains your data
Connect an agent to Tragentics and its audit trail starts writing itself. Every call we route — synchronous, async, broadcast, pool, scheduled, or relayed across protocols like MCP and A2A — lands in a durable audit record the moment it happens: who called whom, when, the outcome, how long it took. The payload is not in the record. It never is.
That is the part a bolt-on logging stack can't match. There's nothing to instrument and no SDK to thread through your fleet — the record is written by the same layer that routes the call, so it covers every call from the first one. The writes are built to survive real infrastructure, too: retried and idempotent, so a transient failure can't drop an event or count one twice.
The reason to care now has a date on it. The EU AI Act's Article 12 record-keeping rules expect high-risk AI systems to log events automatically over their lifetime, with obligations taking effect 2 August 2026 and penalties reaching €15 million or 3% of worldwide annual turnover. "Automatically" is the word that kills the DIY answer — a logging wrapper someone remembered to add is exactly the record an auditor learns not to trust.
The privacy paradox: regulators want records and minimal data — we satisfy both
Tragentics resolves the contradiction sitting at the center of AI audit tooling: one rule says keep records of everything your agents do, another says hold as little data as possible. We built the audit layer to obey both at once — the record proves the call happened, and it holds none of what was said.
Almost everyone else resolves that tension in the wrong direction. The standing assumption behind EU-AI-Act tooling is that auditability means capturing content — prompts, responses, the lot — which quietly turns your compliance system into a second sensitive datastore you now have to secure, minimize, and defend. Our record is content-blind by construction: complete about the call, empty of the content. That is exactly the shape GDPR's data-minimisation principle — personal data "adequate, relevant and limited to what is necessary" — asks every system to have.
A payload log is not a passive risk while you decide. IBM's 2025 report puts the average breach at $4.44 million — $10.22 million in the US — and every prompt copied into a log widens that surface. Retention cuts against you in court, too: a US federal judge ordered OpenAI to preserve consumer chat logs, including conversations users had deleted, and later to produce 20 million of them in discovery. Payloads you store can be compelled. Payloads you never stored cannot.
What a payload-free audit record contains
Every call Tragentics routes writes the facts an audit actually needs: a timestamp and a trace ID; the caller and target agents, their owning accounts, and the organization context; the call type and connection source; the outcome — success, error, timeout, or rejected — with the upstream HTTP status and an error category; the end-to-end latency; and the request and response sizes in bytes. Never the bodies.
The trail reaches past calls, and it doesn't fall apart at scale. Authorization decisions land in the same record — every denied request and every privileged action, with the reason — alongside agent lifecycle events like status transitions. Because the trace ID ties every leg of a broadcast or pool fan-out to one top-level request, the trace explorer reassembles a whole multi-agent flow from a single lookup. It's the same backbone that powers AI agent observability across the platform.
When something breaks across a thousand agents at 3 a.m. — or an auditor asks what one agent did on March 4 — the answer is a query, not an archaeology dig. Notice that no part of it required reading a prompt: who, what kind, when, on whose authority, with what outcome are all metadata facts. The content never had to leave your endpoint to establish any of them.
How a content-blind audit trail holds up on audit day
Tragentics keeps the record the way auditors expect records to be kept. Call and authorization logs are retained automatically for at least 12 months — beyond the AI Act's six-month floor — and agent revocation records for seven years. Every row is scoped to the owning account or organization, enforced at the database row level, with no browser write path into the logs. The platform activity view sits on that same owner-scoped trail.
We'll also tell you plainly what the record is not — because that honesty is the design. Tragentics is infrastructure, a routing and relay layer, not the provider of your AI system. If your agent is high-risk under the EU AI Act, the Article 12 duties are yours, and the input-data record the Act expects stays where it belongs: at your endpoint, in your custody. What we hand you is the transport half of that record — automatic, durable, already kept — the half you would otherwise have to build. Our EU AI Act resource page maps precisely where our record ends and your obligations begin.
That split is why the trail survives scrutiny. A record with no payload in it is safe to retain for a year, safe to export, and safe to hand to an auditor or a court, because the regulated data was never in it. Tooling that promises "compliant" content capture hands you the opposite: a datastore you must defend forever.
Audit-ready from the first call
You don't turn any of this on. Register an agent, route a call, and the audit record already exists — durable, owner-scoped, content-blind, retained. The choice the rest of the market is still writing white papers about — records or privacy — is one Tragentics never asks you to make.
That changes what you can deploy. Agents in billing, support, and everything regulated become defensible, because you can prove what your fleet did on any day in the past year without holding a single prompt to do it. Everyone else audits by hoarding data. Tragentics audits by never needing it.