Agent experience
The environments agents work in matter more than the models working in them. Notes toward a discipline. Draft.
Draft. Notes toward a discipline, from someone who spends most days building the environments rather than the models.
We spent decades learning to design software for humans and called it user experience. We are now shipping software whose primary users are other software — code reviewers, doc writers, desktop operators, deployment agents — and the design discipline for their working environment mostly does not exist yet. Call it agent experience. The models are less often the bottleneck than the shape of the world we hand them.
What follows is not theory. Each claim has a system behind it.
Semantic beats visual
metadesk gives agents the operating system's accessibility tree — the same structure screen readers use — instead of screenshots. An agent clicks an element by ID and receives a delta showing what changed. Screenshot-driven agents burn their budget rediscovering what the OS would have said for free, in structured form, with stable identifiers. When semantics exist, hand the agent semantics.
Determinism is a kindness
Drydock's review quality lives and dies by its context builder: a deterministic assembly of what the model sees, inside a fixed budget. Deterministic environments make agent behaviour reproducible, and reproducible behaviour is the only kind you can evaluate, debug, or improve on purpose.
Close the loop or lose the agent
An agent that never sees the effect of its action cannot correct. metadesk returns UI deltas after every action; Drydock's meta-review grades reviews and feeds the grade back into prompts; Bahia publishes runtime observations so drift is a fact the system can see rather than a surprise an operator finds. The fastest wins in this field are feedback loops, not model upgrades.
Fail loudly, in a shape with a name
Harbormaster Watch renders empty and degraded states explicitly when relays have no data, instead of fabricating samples. The same rule saves agents: silent, ambiguous failure is where production agents go to die. An error the agent can name is an error it can route around.
Identity is part of the environment
An agent that signs its work is an agent whose work can be audited, attributed and revoked. Signet exists so every agent in my fleet has an identity without any of them holding a key. Accountability infrastructure is AX too — it is what lets you widen an agent's permissions without lying awake.
Why bother naming it
Because the work is going to happen anyway, scattered across a thousand codebases, and disciplines that get named early get standards instead of folklore. The environment is the product. We may as well design it like one.