Human operatorVisual layer
Reads the interface.
Hierarchy, labels, proximity, feedback and familiar interaction patterns build a mental model.
01 / AI Laboratory · Agentability
Live public research · agentability.io ↗
Software hasa new operator.
Agentability asks what software must become when an AI agent—not only a person—needs to understand it and act.
I built a public audit, an eight-principle framework and a 100-product Index to turn that question into something teams can inspect. Run the public audit ↗01
The dual-operator surface
One system.
Two ways in.
A person reads hierarchy, scans labels and clicks visible controls. An agent parses structure, identifies state and calls an action. Both need to arrive at the same intent and consequence.
Same software
Intent
state
action
consequence
Agent operatorMachine layer
Parses the system.
Semantics, addressable actions, queryable state and explicit outcomes make operation reliable.
02
The proposed discipline
From human factors
to agent factors.
Human Factors Engineering asks what a person needs to perceive, understand and operate a system. Agent Factors Engineering asks an adjacent question for a different operator: what does an agent need to parse, decide and act reliably?
The distinction matters because human readability and machine readability overlap, but they are not the same. A page can be clear to a person and ambiguous to an agent. It can also be structurally legible to a machine while remaining miserable for a person.
AFE is a proposed framework and ongoing research program. It is not an established industry standard. The public tool makes the proposition inspectable rather than asking the field to accept it on assertion.
03
The v0 audit
Eight factors.
One operating question.
Can an agent understand what is here, identify what it can do, know what happened, and hand control back when it should?
Machine readability
Can the agent reliably parse the page and its meaning?
Chunking
Is information divided into discrete, addressable units?
Control
Can the agent identify and trigger the intended action?
Status
Does the system expose state beyond visual feedback alone?
Defaults
Are starting states safe and usable without hidden context?
Handoffs
Are transitions between tools, agents and people explicit?
Shadow UI
Is there a machine-readable surface beside the visual one?
Transparency
Are intent, logic and consequence visible to both operators?
04
The public instrument
URL in.
Evidence out.
The audit keeps the process legible: render the real page, inspect it against eight factors, then return a report a person or agent can use.
See the real page
A browser loads the interface and extracts its operable structure.
Test eight factors
Structural checks and grounded model judgment identify where operation succeeds or breaks.
Return the evidence
One score, principle-level findings and fixes—available as a readable report and structured data.
05
The evidence base
The framework needed
something to push against.
So I audited the home, pricing and documentation surfaces of 100 SaaS products. The Index is a baseline, not a leaderboard for its own sake: it shows where agent operation breaks across real software.
The live v0 dataset averages 39.7 out of 100. No indexed product crossed 55. The useful conclusion is not that software teams failed. It is that most software was never designed for this operator.
100Products
300Pages audited
39.7Average / 100
<55Highest band in Index
06
The evidence boundary
A score is a prompt for inspection—not a certification.
The current rubric is v0. It weights eight factors equally, and qualitative checks can move between model runs. Content pages can also be penalised by principles designed around application behavior. Reliability testing and confidence intervals are part of the next rubric, so the honest unit is the finding and its context—not false precision in a single number.
Design the interface.
Design the machine surface.
Software now has to explain itself twice: once to the person deciding what should happen, and once to the agent trying to make it happen.
