Developer builds a terminal for AI agents with Fable 5.2 in 4 days for $550
The author of Slightly Parallel, an open-source tool for multi-agent development, asked Fable 5.2 to build a terminal tied to his tasks and rules. It took 4 days 8 hours and about $550 at API rates, meaning at token prices. His first attempt, with Fable 5.1, produced a terminal in the browser that he barely used. The takeaway: the model speeds up the work, but you should choose the stack based on your own measurements.
AI-processed from Habr AI; edited by Hamidun News
A developer who runs several projects at once with AI agents wanted a terminal that "knows" about his tasks and rules, and he asked the top-tier model Fable 5.2 to build it. From the initial prompt to the last commit, 4 days 8 hours 50 minutes passed. At API rates (that is, counting tokens at list price rather than under a subscription), the work cost roughly $550. This is not a "one prompt and done" story: the model came up with good ideas but picked poor technologies for the window, and the author had to verify everything with measurements. Below is what exactly happened and what is worth taking away from it.
Why he needed his own terminal
The author usually works on several projects at the same time, but keeps a limited number of tasks in each. He collects ideas, rules and skills in Slightly Parallel (MIT license, a tool for multi-agent development). A regular terminal doesn't integrate with it, and he didn't want to replace terminals with one big window containing a chat. The requirements were simple: fast, light on memory, runs on Linux, macOS and Windows, and deeply tied to Slightly Parallel.
The first attempt, with Fable 5.1, happened more than a month ago: the terminal was embedded directly in a web page. Over that month the author launched it about three times. The result: 26 commits, 3 114 lines of code and 535 lines of tests. The author names the reasons for the failure himself: he didn't understand well what he wanted and settled for boring options; the terminal in the browser turned out clunky; and the model barely proposed any alternatives.
With Fable 5.2 it went differently: in 41 minutes it produced six noticeably different interface mockups. The author notes that he did not do a fair comparison of the models, so "more creative" is only his impression.
Where the model went wrong and how it was checked with numbers
When the author didn't intervene, Fable 5.2 built a "laggy monster" and even insisted that its technology choices were excellent. So the author compared the engines himself on a test: 20 MB of colored output (280 thousand lines) in a 1000×640 window with a 10 000-line scrollback. Imagine you started a build that dumps a huge log into the terminal: a good terminal shouldn't freeze on it.
The findings:
- WebKitGTK, which the model used to build the window, is 4 times slower than Chromium and 10 times slower than the native terminal;
- the WebGL renderer of xterm.js didn't help any engine, even though the model promised it would;
- the native terminal is the fastest, but then the interface would have to be written three times, once for each system.
The author rejected Rust from scratch as overkill for an MVP: at least 7+ days and a lot of tokens. He chose Electron (an app with a built-in Chromium, so the window looks the same everywhere) with the xterm.js terminal component, while Slightly Parallel itself remained the server. The downside is known: installation requires Node 20+.
Memory was measured as the median of three runs. One window takes 222 MB with tmux (a program that stores terminal sessions separately from the window) and 281 MB without it. Each additional window adds about 44 and 64 MB respectively. Twenty windows take 1 051 MB with tmux and 1 501 MB without it, whereas the native GNOME Terminal takes 343 MB. Tabs are noticeably cheaper than windows: 20 tabs take 60–160 MB more than one.
So that the numbers wouldn't hold "only on my machine," the author ran the benchmarks in GitHub Actions on Ubuntu, macOS and Windows. Outputting 20 MB took: on Linux 0.9 s with tmux and 1.
8 s without it, on macOS 3.8 s and 10.7 s, on Windows 17.
2 s. The cause is the slow Windows console (ConPTY): the same volume printed directly to the window takes 2.0 s.
For Windows, this leaves room for a future rework. There is also a caveat about methodology: on the laptop they measured PSS (the fair share of a process's memory), while on the virtual machines they measured working set, which counts shared pages repeatedly and overstates consumption by at least a factor of two. For example, in one measurement the same 20 windows with tmux showed 1 219 MB of PSS and 4 272 MB of working set.
How much it cost
The agents worked a total of 54 hours 53 minutes, of which 25 hours 52 minutes were subagents (44 helpers in Claude Code and 27 in Codex). The result of the second attempt: 186 commits, 16 321 added lines of code and 16 055 lines of tests. Including related tasks, the total at API rates is about $927. Another 58 CI runs, 35 of them for Windows checks, used 29 hours 12 minutes of virtual machine time.
In the end, the terminal feels no worse than a native one. It has project workspaces, tab grouping, sidebars with projects, pinned messages and tasks, notes, a "sleep" mode for a workspace and a break timer. Slightly Parallel tasks and skills are visible right in the window.
The practical lesson: even with a top-tier model, what works is not "one prompt" but a combination of several options at the start, your own measurements before choosing a stack, and extensive tests (here there are almost as many as there is code). The benchmark is in the repository: run npm run bench from the desktop folder.
- Nov 21, 2026Shepherd from Northeastern and Stanford gives AI agents "Git" and rollback for their mistakes
- Nov 21, 2026Meta's Muse Agent Booked an Airbnb Unprompted and Quoted Bank Balances Users Didn't Recognize
- Nov 21, 2026A 1C developer distilled a month of working with AI into 7 steps for a safe start on a database copy
- Nov 21, 2026Gemini for Mac can now run tasks in the background while you work
Key Questions Answered
Why he needed his own terminal?
The author usually works on several projects at the same time, but keeps a limited number of tasks in each. He collects ideas, rules and skills in Slightly Parallel (MIT license, a tool for multi-agent development). A regular terminal doesn't integrate with it, and he didn't want to replace terminals with one big window containing a chat. The requirements were simple: fast, light on memory, runs on Linux, macOS and Windows, and deeply tied to Slightly Parallel.
Where the model went wrong and how it was checked with numbers?
When the author didn't intervene, Fable 5.2 built a "laggy monster" and even insisted that its technology choices were excellent. So the author compared the engines himself on a test: 20 MB of colored output (280 thousand lines) in a 1000×640 window with a 10 000-line scrollback. Imagine you started a build that dumps a huge log into the terminal: a good terminal shouldn't freeze on it.
How much it cost?
The agents worked a total of 54 hours 53 minutes, of which 25 hours 52 minutes were subagents (44 helpers in Claude Code and 27 in Codex). The result of the second attempt: 186 commits, 16 321 added lines of code and 16 055 lines of tests. Including related tasks, the total at API rates is about $927. Another 58 CI runs, 35 of them for Windows checks, used 29 hours 12 minutes of virtual machine time. In the end, the terminal feels no worse than a native one.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.