1 of 14

ONWARD ACCESSIBILITY · 2026

POUR for Agents.

Context engineering for web accessibility workflows.

Dylan Isaac

Founder, Enablement Engineering

BUFFALO, NY · ENABLEMENT.ENGINEERING

2 of 14

01 · WHAT AN AGENT IS

A language model wired up to tools.

Reads files. Inspects code. Runs tests. Calls APIs. Parses whatever the environment hands back.

STRENGTH

Fast.

Synthesizes across long inputs at a speed humans can't match.

STRENGTH

Text-native.

Code, docs, tests, terminal output — all the same medium.

WEAKNESS

Forgetful.

No durable memory unless the environment hands it back.

WEAKNESS

Pattern-hungry.

Overfits to whichever pattern is nearest. Sounds confident early.

POUR FOR AGENTS · DYLAN ISAAC

02

3 of 14

02 · THE DEFAULT FAILURE

It looks finished.

Underneath, the semantics are thin.

The model learns to make an interface look done before it reliably learns to make it work — for keyboard, for assistive tech, for users.

CAPABILITIES ARE JAGGED.

POUR FOR AGENTS · DYLAN ISAAC

03

4 of 14

03 · THE MEMORY MISMATCH

Humans accumulate.

Agents restart.

HUMANS

Accumulate context.

Past bugs. Deprecated components. The time axe passed and the screen reader experience was still broken.

AGENTS

Restart from zero.

Helpful, fast, tireless — with their own training-data baggage and a fresh-contractor-on-day-one vibe.

Institutional knowledge is invisible to the agent until it's encoded.

POUR FOR AGENTS · DYLAN ISAAC

04

5 of 14

04 · POUR, AGAIN

We make the web POUR

for users.

We need to make our workflows

POUR for agents.

POUR FOR AGENTS · DYLAN ISAAC

05

6 of 14

05 · THE INVENTORY

Inventory the surfaces the agent reads.

01

Code

Component names. File structure. JSX.

02

Docs

READMEs, pattern docs, ADRs.

03

Tests

Names that describe intent, not just status.

04

File names

Dialog.tsx teaches differently than Popper.tsx.

05

Component APIs

Props that name the contract.

06

Tool output

Logs, terminal, axe results.

07

Review notes

PR comments, decision records.

08

Schemas

The shape of correct output.

Every readable surface is an affordance — or, if we ignore it, an obstacle.

POUR FOR AGENTS · DYLAN ISAAC

06

7 of 14

06 · THE FRAMEWORK

POUR, turned around.

P

Perceivable

Can it see the right signals?

O

Operable

Can it act through safe handles?

U

Understandable

Can it follow our reasoning?

R

Robust

Can it check and correct its work?

POUR FOR AGENTS · DYLAN ISAAC

07

8 of 14

07 · WORKED EXAMPLE - W3C Alt Text Decision Tree

Alt text is a decision, not a caption.

NAIVE

image

caption

Pixels alone get you a description of pixels.

WORKFLOW

image

+

context

+

decision tree

alt behavior

The process reveals the information required to make the decision.

POUR FOR AGENTS · DYLAN ISAAC

08

9 of 14

08 · PERCEIVABLE, FOR ALT TEXT

The image alone isn't enough context.

image

+

DOM

+

nearby text

+

page purpose

+

interaction role

Better signals, not more vibes. Pixels plus context get you closer to an accessibility decision.

POUR FOR AGENTS · DYLAN ISAAC

09

10 of 14

09 · OPERABLE, FOR ALT TEXT

Give the agent a handle, not a vibes prompt.

01

Classify

Run the image through the decision tree.

02

Recommend

Decorative, functional, informative, complex.

03

Explain

Show the decision path, not just the verdict.

04

Flag

Route ambiguous cases to a human.

The handle encodes the workflow.

POUR FOR AGENTS · DYLAN ISAAC

10

11 of 14

10 · UNDERSTANDABLE, FOR ALT TEXT

Decision trees are agent context.

a11y/patterns/alt-text.md

Q1.

Does the image contain text?

Q2.

Is it functional — the only label inside a control?

Q3.

Does it carry meaning relative to surrounding content?

Q4.

Is it complex enough that no short alt will do?

Q5.

Is it decorative — redundant with nearby content?

The W3C version is public. What your team knows about your product is not.

POUR FOR AGENTS · DYLAN ISAAC

11

12 of 14

11 · ROBUST, FOR ALT TEXT

The answer must be reviewable.

01

Decision path

The why, not just the what.

02

Tool checks

Schema, presence, structure.

03

Judge review

Internal consistency, not user impact.

04

Human escalation

When ambiguity matters most.

Different accessibility tasks need different loops. Designing the loop for the actual task is part of the work.

POUR FOR AGENTS · DYLAN ISAAC

12

13 of 14

12 · THE TAKEAWAY

Your expertise is the asset.

01

Inventory it.

02

Encode it.

03

Surface it.

04

Verify it.

So agents magnify expertise, instead of generating slop.

POUR FOR AGENTS · DYLAN ISAAC

13

14 of 14

END · Q&A

Thank you.