The Stranger Test · July 2026
How do you treat
your AI agents?
We give every human hire eight gifts on day one — context, memory, scoped access, review, records, a hand near the wheel, a defined role, and a fair hearing when things break. This test scores whether your agents get any of them. Eight questions, about 90 seconds, graded 0–16. No email to take it — none to see your results.
The First Week · the Stranger Test
You just hired a brilliant stranger. It’s Monday.
Eight decisions across their first week — about 90 seconds. No email to play, none to see your result. You’ll leave with your Management Style Card.
What the Stranger Test measures
An AI agent is a brilliant stranger in your house: enormous capability, zero knowledge of your world. Organizations already know how to absorb such a person — the Stranger Test simply asks whether your agents get the same treatment, across the eight dimensions of management we consider non-negotiable for any human hire. Published July 2026; the questions score your current setup, not your intentions.
- 01 Context
- Do your agents start knowing the project — conventions, architecture, landmines — or interrogate it cold every session?
- 02 Memory
- Does anything an agent learns persist to the next session — and reach the other agents?
- 03 Permissions
- Are agents scoped to a role, or do they run with your full access?
- 04 Review
- Is risky agent output checked before it lands — by something other than another LLM’s opinion?
- 05 Audit
- Could you replay exactly what an agent did last Tuesday — from records, not memory?
- 06 Intervention
- Can you redirect or take over a running agent mid-task — without killing the run?
- 07 Structure
- Are your agents specialists with defined niches, or one generalist prompt for everything?
- 08 Accountability
- When an agent fails, can you tell why from records — or only guess?
Scoring
Each dimension is answered on three tiers — systematic (2 points), partial or manual (1 point), absent (0 points) — for a total of 0–16. Four grades:
- 13–16Managed fleetContext, memory, scope, review, records, a hand on the wheel.
- 9–12Supervised chaosGood instincts, honor-system enforcement.
- 5–8Keys to a strangerEnormous capability, nearly zero management.
- 0–4Root access and a prayerRoot and no memory. You’d never do this to an intern.
The eight dimensions are the same ones the manifesto argues we already honor for people, and the same ones AI agent management exists to systematize. If your grade stung a little —that’s what we’re building.