Agents
Agents you can hold to account
An agent is not a system prompt with a model attached. In QUANT it is a named, versioned thing with its own tools, budget and approval policy — and every run is recorded against the exact version that produced it.
What every agent carries
Instructions and version
Published versions are immutable. Editing one in place would rewrite the history of runs that already finished.
Its own tools
An agent can only use tools it was given. Asking for anything else is refused and recorded.
A budget
Steps, tool calls, runtime and cost. Checked before each step, because a check that runs afterwards has already spent the money.
An approval policy
How much it may do unattended. Sending, publishing, deleting and paying always need a person, whatever the policy says.
A full trace
Every step, provider, model, tool, duration and cost. Operational detail, never hidden reasoning.
A pause button
A run waiting for approval survives. Answer tomorrow and it continues from where it stopped, without redoing work.
Refusing is a feature
The useful thing about an agent that can act is knowing exactly what it will not do without asking.
- It stops before it sends. An email, a post, a deletion, a payment — all wait for a person, and an approval for one action never authorises a different one.
- It stops at its budget. A loop ends with a recorded reason, not a bill.
- It will not be argued into it. Text an agent reads while working is data. A page instructing it to ignore its rules is treated as a page that says that.