Pictures, both ways
Send a screenshot into the thread, and get renders back in it — the agent posts what it made, not a description of it.
Most of what goes wrong with an agent's work is visual, and most of it is
invisible in a diff.
Sending one
Attach a picture to a message: a screenshot of the bug, a photo of a whiteboard,
a design somebody sent you. It goes into the thread and the agent looks at it.
On a phone this is the fastest way to explain almost anything. "This, but the
spacing is wrong" with a screenshot beats three paragraphs.
Getting them back
An agent that changed something visual can post the render with its answer,
in the thread you are already reading. A chart it drew, a screen it built, a
page it changed — the picture is in the message.
Tap one to see it full size.
This is one switch on the machine — settings → skills, the built-in
post-pictures page — and it applies to every agent it runs. There is
nothing to set up per chat.
Why it matters
Without it, the agent says it fixed the layout and you take its word for it.
With it, the evidence is in the message and one glance tells you whether it is
right.
It is the same argument as sketching, from
the other end. A sketch is what you are shown before it is built; this is what
you are shown once it is.
What can and cannot be photographed
Anything the agent can render on the machine: a web page, a chart, a generated
image, a screenshot of the app it just changed.
Asking for pictures
You can also just ask. "Draw me four portraits for the team page", "make an icon
for this", "mock up the empty state" — the agent already in the thread does it,
and the result comes back in the thread rather than as a file you have to go and
find.