How do you test whether an AI agent is durable?

Take one agent your company is running or planning and score it on each of the five properties, zero to ten. The anchors under each property describe what a two, a five and an eight look like in production.

What does the durability score mean?

7.0 and above: This system can probably survive production. 4.1 to 6.9: This system is at risk. 4.0 and below: This system is likely to fail.

What score does each property need to reach?

Recoverable: Below seven is a major risk. Resumable: Below seven and you are losing work. Auditable: Below five is a compliance risk. The bar is low because almost nobody clears it today. That is a statement about the industry, not about how much it matters. Replaceable: Below six and you are coupled too tightly. Coordinated: Below six is a coordination problem.

What evidence should back each score?

Recoverable: Show the last incident where a worker died mid-workflow and what the agent did next. Resumable: Show a workflow that paused for a person for more than a day and completed correctly. Auditable: Pick one decision from last month and produce its record now. Replaceable: Name the last model change and how long it took to reach production. Coordinated: Show two agents that touched the same record and how the system prevented a conflict.

The Durability Score

Score your agent

Take one agent your company is running or planning. Score it on each property, zero to ten. The anchors under each one describe what a two, a five and an eight look like in production. Be honest. Nothing here is stored. Copy the result or share the link if you want to keep it.

What the number means

7.0 and above: This system can probably survive production.
4.1 to 6.9: This system is at risk.
4.0 and below: This system is likely to fail.

01
Recoverable
If the system fails, does it detect the failure, pause, and resume when the failure is fixed? Or does it crash and leave state inconsistent?
2
A crash means starting the workflow again from the beginning. Someone checks by hand what already happened.
5
Steps are checkpointed. Resume works most of the time. Some retries duplicate side effects.
8
Every step resumes exactly where it stopped with prior outputs intact. No duplicate side effects. Verified under fault injection.
Show the last incident where a worker died mid-workflow and what the agent did next.
Below seven is a major risk.
02
Resumable
If the agent pauses mid-decision, can it resume exactly where it left off without repeating work or side effects? Or does it restart?
2
There is no pause. The agent completes or fails. Human review happens after the fact.
5
The agent can wait for approval, but the wait has a short limit, and long waits time out or restart.
8
The agent waits days if it must, the wait is an audited event, and it resumes from the exact point with all state intact.
Show a workflow that paused for a person for more than a day and completed correctly.
Below seven and you are losing work.
03
Auditable
Can you show a regulator or a customer the complete decision trail for one transaction: what was decided, why, on what data, by which model version, whether it was reviewed or overridden, and what the outcome was?
2
Application logs. Timestamps of activity. No record of inputs or reasoning.
5
Prompts and model outputs are logged. Inputs are partially captured. Human interventions are not linked to decisions.
8
One query returns the full decision record: inputs, model and policy versions, every tool call, every intervention, confidence, and anything proposed and refused.
Pick one decision from last month and produce its record now.
Below five is a compliance risk. The bar is low because almost nobody clears it today. That is a statement about the industry, not about how much it matters.
04
Replaceable
Can you change the model without rebuilding the workflow? And if the agent had to come out of service, would the business process continue without it?
2
Prompts, tool bindings and error handling are written for one model version. Changing it means revalidating everything.
5
The model is behind an interface. A swap needs prompt rework and a full regression pass.
8
Model and workflow are separated by design. A swap is configuration and a targeted test. Two models can be compared side by side.
Name the last model change and how long it took to reach production.
Below six and you are coupled too tightly.
05
Coordinated
Are the handoffs between the agent and other systems explicit, documented and auditable? Or is coordination happening through logs that are hard to parse?
2
Agents call each other directly. Failures between them are discovered by reading logs.
5
Handoffs go through a queue or an orchestrator. Ordering and conflicts are handled case by case.
8
Every handoff is a durable, ordered, auditable event. Shared resources are locked. Disagreements route to a person by design.
Show two agents that touched the same record and how the system prevented a conflict.
Below six is a coordination problem.

0 of 5 set. The result appears when all five are scored.

This is the first of the three Monday-morning moves in the book. The second is to take the agent that would hurt most if it failed and rebuild it durable. The third is to find the person in your organisation who will hold the bar.

Thresholds are the book’s. If you want a second opinion on the score, Technoidentity runs an assessment of one agent workflow against these properties. You keep the report. The assessment.