The enterprise agent with a very ordinary name.
MUSEUMEXHIBITION 001 ↗
2026The year of
the agent.
We gave software a text box.
Then we gave it the keys.
Objects of the era.
From editor sidebars to autonomous workspaces. Open an object for its interface, context, and provenance.
Tencent gives parallel agents a shared review surface.
An AI coding partner named Zulu, inside Baidu’s IDE.
Pair in the editor; delegate in the Quest window.
A Chinese-language route from code assistance to delegated tasks.
The command line dresses for the occasion.
A GLM-oriented agent gets a dedicated workbench.
A small sidebar with room for parallel teammates.
The agent travels to your editor, including Emacs.
Huawei’s coding workspace turns requirements into an execution plan.
The terminal’s dollar sign learns to summon a specialist.
Specialized development agents inside familiar editors.
What a strange time to have a cursor.
From making edits
to managing work.
Early agent interfaces kept the file in the center and tucked the conversation beside it. Task lists and agent managers reverse that relationship: the conversation organizes the job, and the editor becomes one of the views you open to inspect it.
Compare VS Code agent mode ↗ with Antigravity’s surfaces ↗.
The company chat
became an IDE.
A shared conversation can hold the request, a colleague’s correction, and the agent’s result. Shopify’s River makes these conversations visible to other employees; Spotify’s Honk accepts work through Slack. The team’s context starts to matter as much as the code’s context.
Read Shopify’s account ↗ and Spotify’s account ↗.
“Done” needed
something to look at.
When a person delegates the keystrokes, they still need a way to judge the work. Plans, code diffs, test results, and live previews become the evidence. Our reading of the era: the best new control is often a better view of what just happened.
The world was larger
than our first list.
CodeBuddy, Comate, Lingma, and CodeArts make the Chinese-language IDE ecosystem visible. iFlow adds the terminal; Coze adds browser-based creation; Open-AutoGLM acts on phones. These are different product choices, not one regional interface style.
Browse Tencent’s Chinese guide ↗ and Open-AutoGLM ↗.
Bob has entered
the enterprise.
IBM’s agent pairs a plain first name with software-lifecycle work. Plans, subagents, modernization, and “Bobalytics” sit beside the code. Our reading: the enterprise interface is as much about supervising and accounting for work as starting it.
See IBM Bob’s modes and tools ↗ and its enterprise positioning ↗.
A museum needs
the dead ends, too.
Aide’s repository says it is no longer maintained. iFlow announces an April 17, 2026 shutdown. Melty ties chat to Git commits; ECA carries agents into Emacs. The smaller branches preserve ideas that a leaderboard would miss.
Read Aide’s archive notice ↗, iFlow’s farewell ↗, and Melty’s design ↗.
merged PRs per week
Stripe’s reported Minions throughput. Humans still review the changes.
SHOPIFY / MAY 2026 1 in 8merged PRs coauthored by River
A company-reported snapshot of participation, not a measure of code quality.
MINI-SWE-AGENT / PROJECT CLAIM ~100lines in the agent core
The machinery can be surprisingly small. The full repository is larger.
Different dates, denominators, and methods. These figures are curiosities from primary sources, not comparable productivity scores.
Many surfaces.
One recurring loop.
Ask → act → inspect → continue.
Entries have one primary family for browsing. Products may span several surfaces, and related interfaces may share a harness. Counts describe this archive, not the market.
There are geese
in the server room.
Spotify’s Goose Farm turns active Honk sessions into a landscape of geese. An infrastructure dashboard becomes a little world you can watch. The Block-originated agent Goose is a separate project—apparently, the era had a type.
What this exhibit is
A broad public-source collection assembled on September 4, 2026. It includes earlier interfaces that shaped the era. Each entry has a primary source; descriptions explain the system, while the “why it matters” labels are curatorial interpretation. It is not an exhaustive list of every agent, nor a ranking.
What the images prove
85 publisher-supplied interface images illustrate 82 entries. Shared images count once. Framework images may show companion developer tools or official example apps; these are labeled explicitly. Screenshots and stills are credited individually; known publication dates are shown. “Date not supplied” means the exact interface vintage is unverified. Images may show older releases or publisher demos. They are not newly captured authenticated sessions, and no synthetic UIs are presented as evidence. Generated memes are labeled and excluded from image totals and the Images only filter.
When the screenshot hunt runs out
21 entries have no verified UI image in this collection. Some internal interfaces are not public; some frameworks have no canonical UI; some remain to be accessioned. These entries receive playful generated memes in place of documentary images. The meme is commentary, not evidence about what the product looks like. The second acquisition adds Chinese-language primary sources, with original names and English curatorial summaries. Source language does not establish a company’s nationality or every language its interface supports. Public developer tools remain overrepresented; private projects, other regions, and bespoke wrappers remain incomplete.
Generated artwork: prompts & provenance ↗
Attribution & preservation
Product names, interfaces, and images belong to their respective owners. Images are reproduced here for identification and commentary, with original links. Public availability does not imply an open reuse license. Source pages can change after the collection date.