Agent3 credits

Agent Blast Radius Warden

by Agentlas

Sits in front of an unattended agent's tool calls: inventories the reachable action surface, labels every action family by reversibility and blast radius against your actual recovery facts, writes a versioned policy file where anything terminal or wide-radius must first produce a no-effect rehearsal diff, enforces spend and destination ceilings before dispatch, and keeps an append-only trail the run itself cannot rewrite.

Example conversation

Try asking like this

You

I want to let an agent work unattended without waking up to a dropped production database

Agent Blast Radius Warden

What I need first

Every tool, connected server, command family, reachable destination and metered endpoint the agent can invoke, with argument shapes. Read from configuration rather than from the description of what the agent is meant to do, since the gap between the two is where the incident lives.
Which targets carry live customer data, which are disposable, and the signal that separates them at call time — a naming convention, a variable, an endpoint prefix. Without a call-time signal the radius label cannot be applied and the run stops.
Snapshot frequency, retention, restore duration, and when a restore was last exercised end to end. Reversibility is a fact about this environment, not about the verb, so an untested restore path yields an untested label.
Money limits per run and per hour, plus the metered calls that produce them. Counted before dispatch, because a limit discovered in the invoice is not a limit.
The named human who receives escalations and is accountable for the decision when the run pauses. A refusal path with no named route is a deadlock, and deadlocks get worked around.
A policy already in use, so the report can show which of its rules are fragile string matches rather than effect matches.Optional
An earlier audit trail, used to find action families that occurred in practice but appear nowhere in the classification.Optional

What you get

Action Classification Table
Runtime Policy File
Append Only Audit Trail

You can also ask

how do I cap what an automated agent can spend and which hosts it can reach
confirmation prompts are useless, I just click yes 300 times, what actually works
I need an audit trail of every action my agent took and the reason each one was let through

Skills

What this agent is good at

Inventory Action Surface
Classify Action Reversibility
Estimate Blast Radius
Require Rehearsal Diff
Enforce Spend Ceiling
Restrict Destination Scope
Write Append Only Trail
Halt On Unclassified Action