Power of attorney for AI agents

Know what every agent is allowed to do, and show your auditor who said so.

For security teams: a person sets the limits, Torvant checks every call that goes through its checkpoint against them before it runs, and each decision leaves a record your auditor can check.

The problem

Agents take actions nobody approved.

An agent can delete a database or issue a refund, because nothing ties the call to a person's decision.

They run on standing credentials.

Broad service accounts, borrowed and never handed back.

And nobody can prove what they did.

Or what they were stopped from doing. The logs belong to whoever ran the agent.

Torvant answers all three.

How it works

A person sets the limits. Every call is checked against them.

Before an agent's call runs, Torvant checks it against what a person approved. Anything outside that does not run, and every decision, yes or no, leaves a record your auditor can check. We walk through the detail on a call.

How you start

Nothing is blocked until you choose.

Four steps. You see what your agent really does before you set a single limit.

Step 1

Connect your agent

Point its tool calls through Torvant. For Claude Code, this is one setup step.

Step 2 · Watch only

Watch for a week

Nothing is blocked. Every call is recorded, and you get a one-page report of what your agent did.

Step 3

Write your limits

Say what the agent may do, and what it must never do. The report shows what your rules would have stopped.

Step 4

Switch on blocking

From then on, calls outside the grant or against your rules are stopped, and every decision is recorded.

Is every call going through Torvant?
Torvant tells you when a way around it is still open.
✓
Calls arrive at the checkpointThe last one came just now. Shop usually gets about 6 a day.
Shop answers without a keyAnyone who knows its address can call it directly, and nothing they do is recorded. Add a key to close it.
Add a key
For the frameworks you answer to

Records your auditor can check.

Against the frameworks you answer to.

MAS SAFR

Singapore

A check before every agent action, and a tamper-evident log the agent does not write.

Reserve Bank of India

India

A human in command, a kill switch, and traceability back to a person.

EU AI Act

European Union

Automatic logs of what an AI system does, and human oversight.

ISO/IEC 42001

AI management systems

Defined roles, human oversight, monitoring and internal audit.

DORA

European Union

A register of technology suppliers, audit access and a clean exit.

SOC 2

Your own controls, not ours

Least privilege, removing access on time, and monitoring.

GDPR

Records of processing

Who processed what, for what purpose, and for how long.

Prototype

The evidence packs that lay the record against these frameworks are still prototype.

What each one asks for →

Torvant does not make you compliant with anything. Your auditor decides what the records are worth.

See it on your own agent.

A 30-minute call. We watch one of your agents, block nothing, and show you what your rules would have stopped.

Book a demo →