A museum catalogs 103 coding agents from 2026 and confirms the shift: from typing to directing bots

By Zendoric · September 5, 2026.

An independent curatorial project, published on September 4 under the name "Agent Museum", has catalogued 103 AI coding agent tools that defined 2026, with 85 interface screenshots illustrating 82 of those entries. The archive groups them into six families: 28 editors, 18 workspaces, 17 terminals, 15 frameworks, 13 in-house corporate tools and 12 "beyond code" uses. The author himself clarifies that it is neither a ranking nor a market census, but a collection of public sources with original screenshots and links.

What is interesting is not only the figure —103 entries for a very recent phenomenon— but its geography and its shape. Alongside familiar Western names, an entire bloc from China appears: Tencent's CodeBuddy, Baidu's Comate (with an assistant called "Zulu"), Alibaba Cloud's Lingma IDE, Huawei's CodeArts Agent and ZCode from Z.ai, the company behind the GLM models. Add to that iFlow CLI, a terminal which, according to the museum itself, was already discontinued on April 17, 2026. The entry devoted to IBM Bob —an enterprise agent with a deliberately bland name that manages plans, subagents and legacy code modernization— symbolizes these tools' arrival in the most conservative corporate software.

The catalogue also offers some adoption figures reported directly by the companies, which should be read with the caution the archive itself asks for: they are neither comparable to one another nor independently verified. Stripe stated in February 2026 that more than 1,000 pull requests (proposed code changes) were merged per week through its internal "Minions" system, always with human review. Shopify reported in May that one in eight merged pull requests was co-authored by its agent River, integrated directly into the company's chat. And the mini-swe-agent project boasts a core of barely around 100 lines of code, a reminder that the complexity of these systems lies not in the engine but in how they are directed.

The museum's underlying thesis —and here it agrees with something we have been observing for months— is that the center of gravity of working with code has shifted. The first agents lived outside the editor, suggesting lines; now the conversation with the agent organizes the task, and the editor is just one of the windows for inspecting the result. Tools such as VS Code's task managers or Google's Antigravity panels treat code as one more view, not as the central object. In parallel, company chat has become an entry point to development: Spotify accepts assignments for its agent Honk directly via Slack, and has gone so far as to turn its active-sessions panel into a landscape of animated geese —a curiosity that, beyond the joke, reveals how normalized it has become to have agents working in the background, all the time, as part of the office scenery.

Our reading is that this catalogue, though modest in reach —it was born on Hacker News with a single upvote and no comments—, is a useful symptom of two things we had already been pointing out. The first is the speed of these tools' life cycle: there are already documented "dead ends" —Aide unmaintained, iFlow shut down—, a sign of a market saturating and being culled almost in real time, not of a mature experiment. The second is the confirmation, with very concrete examples, of our thesis on tech employment: inside the very companies building these tools, what gains ground is supervision, architecture and judgment about what the agent delivers, and what erodes is the mechanical task of writing the line of code. IBM calls it "Bobalytics"; we call it, plainly, the administrative side of the programming trade running out of work to do.

The massive and simultaneous presence of Tencent, Baidu, Alibaba, Huawei and Z.ai in a catalogue initially conceived in Western terms is another data point worth not overlooking. It does not measure model capability —this is a showcase of interfaces, not a benchmark— but it does confirm that the tooling ecosystem around Chinese generative AI is advancing in parallel with its models, with its own terminal, IDE and workspace products, not just clones. It is the same gap-closing dynamic we documented in models such as GLM or Qwen, now carried over to the product layer.

In the long run, snapshots of this kind —chaotic, with figures that do not square with one another, with products that die a few months after being born— are exactly the texture one would expect from a real transition. The underlying promise remains intact: the more routine tasks of programming, reviewing and deploying code that agents absorb, the more time is freed for people to concentrate on deciding what to build and why. The museum documents the noise of that transition; the signal, if our thesis holds, will arrive when that noise translates into better and cheaper software for everyone, not just into fewer people writing lines of code.