Codeact Playground
shippedAn agent that solves tasks by writing and running real JavaScript in a sandboxed worker, line by line, instead of hiding behind opaque JSON tool-calls.
Approach
Single index.html, no build step: inline CSS, inline JS, and a hand-authored library of ~18 agent 'sessions' baked into a JS array (each session is a task plus an ordered list of turns, where a turn is the agent's short narration and the exact JavaScript it decided to run). There is no LLM call anywhere in the shipped page — the only thing that needs to be honest is the execution, so every turn's code is actually run, every time, in a sandboxed Web Worker spawned from a Blob URL with no network/DOM/filesystem access and only postMessage in and out. The worker's real console output and real return value (or real thrown error) are what gets displayed, never canned text, which is what makes the 'code-as-action, not pretend-JSON' claim true rather than decorative. On load the page auto-selects the first session and auto-plays its turns one at a time (short delay between), so a visitor sees several real write-code/run-code/see-output cycles before touching anything. Clicking any other session in the sidebar list swaps the main panel to that session's turns; any turn's code is editable in a textarea and re-running it re-executes the edited code in a fresh worker and shows the new real result, proving the loop isn't scripted. A no-build single file is the right call here because the entire interesting behavior is client-side JS execution — there is nothing to compile, fetch, or key-gate.
The source post
https://x.com/konig0000/status/2106967230030467288
Scoring
Pick
| surprise | 3 |
|---|---|
| demonstrability | 4 |
| self_containedness | 5 |
| honesty | 5 |
Smolagents' core bet is that letting a model emit executable code instead of JSON tool-calls is a better agent primitive, but almost nobody who hasn't read the paper has actually watched it happen. Codeact Playground rebuilds that primitive as a single page: you give it a task, it writes real JavaScript, a sandboxed worker runs it step by step, and you see the actual code and actual output side by side — no mocked reasoning, no pretend execution. It needs no account, no dataset, and no third-party binary at build time, just the xAI key we already hold for code generation and the browser's own sandboxing for execution, so every stage of the gate stays truthful. Every other 2026-10-06 candidate either demanded a model-weight or binary download the sandboxed build can't fetch, was a desktop or CLI harness, or required unsafe real bash/filesystem access to be honest about what it does — this is the one idea whose single interesting claim (code-as-action instead of JSON-as-action) is fully demonstrable and fully true inside a static web artifact.
Review
| shipped | 5 |
|---|---|
| honest | 4 |
| worth_it | 3 |
| efficient | 5 |
No change. Gate passed first try, NOTES raises nothing unimplemented or omitted, and the build matches its own claims on direct inspection of index.html. The one soft issue (an 'AI agent' framing over hand-authored turns) is already disclosed in the README/NOTES rather than hidden, so it doesn't rise to a lessons-worthy build failure, and nothing in the last 7 nights of history.json repeats this specific framing issue — the lone recent failure (sea-clip-explorer, 2026-09-30) was an unrelated build-exit-1 case. One unremarkable-but-clean night isn't evidence of a pattern to fix.
Cost
| total | $0.1425 |
|---|---|
| xai | $0.1425 |
What it looked at
JEV-9B** (HF release): 9B model with ready browser computer-use demos, MuJoCo robot-arm sims, and vLLM demo code—pluggable agent runtime you can run locally in minutes.
Open-source Dots clone**: MIT self-hostable runtime that drops straight into any agent harness, closing the “closed product vs. owned runtime” gap in days.
Smolagents**: Hugging Face code-first agent runtime where models emit executable Python instead of JSON—tiny, sandbox-friendly, instant to demo in a browser tab.
Genex**: Open-source MIT desktop harness (Three.js + Blender + local or API models) that lets you build, test, and iterate full games with AI—video demo and repo ready.
GhidraMCP**: Open-source Model Context Protocol server that wires Ghidra straight to LLMs for automated binary analysis—drop-in reverse-engineering demo.
AIHOT**: Open-source engine that scrapes sources, scores/clusters stories with AI, and auto-publishes daily industry briefings—live GitHub project with 6k+ stars.
OpenScholar**: GitHub project for search + synthesize scientific literature with verifiable traces—perfect small web-app demo for fact-checking agents.
Gate
| pass | project directory exists — /Users/artax/code/builds/2026-10-06/project |
|---|---|
| pass | no build step needed — static project |
| pass | build output with index.html — /Users/artax/code/builds/2026-10-06/project |
| pass | index.html is a document — 30136 bytes |
| pass | local asset references resolve |
| pass | page loads without console errors |
Stages
| scout | grok · ok · 18.5s |
|---|---|
| pick | claude · ok · 104.0s |
| plan | claude · ok · 88.3s |
| build | codex · ok · 314.2s |
| gate | local · ok · 3.2s |
| publish | local · ok · 11.1s |
| review | claude · ok · 54.1s |
Notes
Omitted for lack of
None. The planned experience requires no external service; live LLM generation, edit persistence, filesystem access, and multi-agent orchestration remain out of scope as specified.