blog · field notes
Blog
Field notes from building audit infrastructure for AI agents.
Product
Models & cost: what every call costs
The same tape that answers your auditor answers your CFO: every model your agents used, the tokens they spent, and what that cost at your vendor — per model, from evidence rather than vibes.
·3 min read
Product
Registry: the agent inventory
Auditors ask it first: show me your inventory. Registry is every agent classified by risk, approved or pending, with a named owner — and a Register PDF you hand over as a document.
·3 min read
Deep dive
Tamper-evidence in practice
Tamper-evident is not tamper-proof — and the difference is the point. How the chain works, where a break actually shows up, and why not crying wolf matters as much as the alarm.
·4 min read
Deep dive
The per-event signature
Eight months from now, someone disputes an agent's action. Most logging stacks can only escalate that argument. A signed event ends it — the quiet workhorse of the whole product, explained.
·3 min read
Product
Anomalies: the 7 rules
At production volume, nobody reads the tape. So Wytness reads it for you: seven rules, per-agent baselines, every alert linked to the signed events that raised it. Here they are in plain language.
·3 min read
Quickstart
Already on AGT? Direct ingest, no SDK swap
AGT's native CloudEvents output points straight at our ingest — zero Wytness code on your runtime. What you get, what needs the wrapper, and why starting simple locks you out of nothing.
·3 min read
Quickstart
TypeScript SDK in 5 minutes
One npm package, the same five env values as Python, one init() — and your Node agent's actions arrive on the tape, signed and sealed. Byte-for-byte the same wire format as Python.
·3 min read
Quickstart
Python SDK in 5 minutes
From pip install to your first verified event on the tape — honestly about five minutes, if your keys are ready. Install, five env values, one init call, watch the row arrive.
·4 min read
Compliance
DSARs when an AI touched the data
A subject access request lands — and this year, agents have been reading and acting on that person's data. "Tell me what you hold about me" now includes what your agents did with it.
·4 min read
Compliance
ISO 27001 when agents are in scope
ISO 27001 audits your ISMS, not your technology. When your organisation deployed AI agents, did your ISMS notice? Your certification auditor will.
·3 min read
Compliance
EU AI Act Article 12: what record-keeping requires
On 2 August 2026 the EU AI Act's high-risk enforcement window opened, and Article 12 turns from paperwork into engineering: automatic logging of events relevant to identifying risks. Decoded.
·3 min read
Compliance
SOC 2 evidence for agent activity
Your AI agents entered your SOC 2 scope the moment they started touching systems your controls govern. Nobody sent a memo — but your auditor will notice. The question is what evidence you'll have.
·3 min read
Perspective
You control the harness, not the model
ASD's new paper tells boards to govern the software around the model, not the model. Of the eleven harness components it names, one is a record of what the body did. That's the one we build. What the paper says, and where a recorder honestly fits.
·7 min read
Deep dive
Pseudonymisation: how we never see your PII
Not 'encrypted on our servers' — it never arrives. How sealed fields keep their analytical shape, how the reveal runs in your browser, and where the boundary honestly sits.
·4 min read
Product
Permissions: what they're allowed to do
Config files say what was granted; they say nothing about what actually happens. Wytness answers from the other side: the permissions your agents actually exercised, action by recorded action.
·4 min read
Product
Tools: what your agents touch
The model never touches your systems — the tools do. If you want to know what your agents can actually affect, don't read their prompts. Read their tools.
·3 min read
Product
Sessions: one run, end to end
A single event tells you what an agent did — rarely why. The refund only makes sense next to the lookup before it and the email after it. Sessions regroup the tape into stories.
·3 min read
Product
Agents: who ran this
After 'what happened' comes 'who did it.' The Agents page is a census of every automated actor in your estate — built from the stream itself, with each agent's track record attached.
·3 min read
Product
Inside one event
Click any row and the full record opens — the panel you'd screen-share in a dispute, and what an auditor actually reads. One refund event, walked top to bottom: who, what, sealed fields, proof.
·4 min read
Product
Events: reading the tape
Every product has one page that is the product. Ours is Events: one row for every action any of your agents has ever taken, in order, each row carrying its own proof. A tour of the tape itself.
·4 min read
Perspective
Observability is not accountability
You already have traces, dashboards, and token counts. Would you hand them to a regulator? Observability answers your questions about your system; accountability answers other people's — and they don't trust you.
·4 min read
Perspective
Why we built on Microsoft AGT instead of our own chain
We had our own signing and hash-chain code in production. When Microsoft open-sourced the Agent Governance Toolkit, we retired ours and rebuilt on the standard. Easiest decision we've made — here's why.
·4 min read
Perspective
You hold the keys
What happens to your data the day your vendor gets breached? Our answer: an attacker would find tokens they cannot reverse and signatures they cannot forge — because the keys that matter were never ours to lose.
·5 min read
Quickstart
Instrumenting a LangChain agent in 5 minutes
Wire the wrapper in once, add one callback, and every tool call your LangChain agent makes lands on the tape — signed, sealed, verifiable. The current five minutes, replacing the retired v1 guide.
·4 min read
Deep dive
How Wytness works
A walk through the full path of one event: from the moment your agent emits it, through signing and buffered ingest, into the verification, hash-chain, and dual-store archive that an auditor pulls back out years later.
·7 min read
Perspective
The blackbox recorder for AI agents
Planes don't carry a flight recorder because crashes are common. They carry one because when something goes wrong, “trust us” is not an acceptable answer. AI agents have reached the same moment.
·3 min read
Perspective
Why AI agents need an audit trail
If a customer support agent issues a refund at 3am, who authorised it? AI agents are taking real actions on real systems, and they need the same audit story human operators have always had.
·3 min read