From notebooks to working apps.
Five notebooks, four build variants: GPT-6 Astra, Qwen on Mac, Qwen on NVIDIA RTX, and Qwen Flash family builds with generation identities stated below. Compare Text Atlas, demand forecasting, energy planning, CPU scaling and Iris classification. Each card identifies the build model. AGILAB presents the working app, its original source bundle, and its recorded checks.
Demo results — Tokki 1.0.76
One actual sensor-quality repair, 16 external checks, a fresh notebook kernel and three deterministic replay objectives. Five RTX nodes accepted 1.0.76; historical strict-doctor warnings remain.
Démonstration du 23 octobre 2026 — Tokki 1.0.76 : résultats et limites ↗
Build from your own notebook
Set up your licensed Tokki installation and your own Codex provider, then install uv.
uv tool install "agilab[notebook-agent] @ git+https://github.com/ThalesGroup/agilab.git"
agilab-notebook-demo --uiChoose Local notebook or Pinned GitHub notebook. Start with self-contained Python and locally available data and dependencies. The local agent uses your configured provider and executes generated Python on your computer. Checks establish execution and interface behavior; review scientific conclusions yourself.
After a successful build, you may choose to share a completion receipt on GitHub. It excludes notebook contents, prompts, paths and credentials. Nothing is submitted automatically; public reports are associated with your GitHub account.
RTX builds use the OpenCode agent with local Qwen; Codex supervises the build and validation online. Recorded build durations, including repairs and checks. One run per build; hardware, quantization and verification scopes differ. RTX timings exclude initial installation, downloads and browser checks. 1 KLOC = 1,000 Python code lines across the generated app and notebook, excluding comments, docstrings, tests and supplied helpers. Lines/min divides that code volume by elapsed build minutes; it does not measure code quality. Build receipts ↗
Verified builds · local execution
Local code generation: Qwen Flash-Next UD-Q4_K_XL on CPU (rtx5). These quantized local variants differ from the remotely served Hugging Face Flash model. Coordinator diagnostic prompt interventions: 9. The other Flash builds retain their recorded remote-provider provenance.
| Stage | Execution location |
|---|---|
| Local code generation | Quantized Qwen Flash-Next on CPU; coordinator diagnostic assistance where recorded. |
| Other code generation | Recorded Hugging Face / OpenRouter routes. |
| App execution and acceptance | Local independent notebook, UI, workflow, numerical and generated-test checks. |
| Public gallery | Sources and measurements; hosting is separate from generation. |
| App | Build models | Active build time |
|---|---|---|
| Iris decision lab | Qwen Flash | 2 min 54 s |
| Text Atlas | Qwen Flash | 4 min 03 s |
| Demand Forecast | Qwen Flash | 3 min 11 s |
| Free-threading lab | Qwen Flash-Next · UD-Q4_K_XL · local CPU | 319 min 11 s |
| MILP Energy Lab | Qwen Flash-Next · UD-Q4_K_XL · local CPU | 343 min 44 s |
These builds use the same original notebook objectives and independent acceptance checks. Each build duration sums generation, repair attempts and notebook, UI, workflow and numerical validation. Environment installation, cached model downloads, operator pauses and deployment checks are excluded. Hardware and workflows differ across editions; these are recorded runs, not a controlled speed benchmark.
Compared with Astra: the published runs use different timing scopes. Astra's historical receipts include waiting time; Flash sums active attempts. With one recorded build per app and different infrastructure, these values do not establish that Flash is faster than Astra.
Build evidence · Download all five appsAstra and Qwen RTX include model turns that drive tools and checks. Qwen Mac and Qwen Flash count file-generation and repair requests; their external orchestration is excluded. Remote Qwen Flash totals also include calls to repair models. Local quantized totals count Qwen generation POST attempts, including interrupted calls; coordinator diagnostic assistance is stated separately. Missing usage stays unknown. These different workflows are not a like-for-like comparison.
| App | AstraAgent loop | Qwen MacFile output | Qwen RTXAgent loop | Qwen FlashFile output |
|---|---|---|---|---|
| Iris decision lab | ≥ 16 | ≥ 3 | ≥ 26 | 4 |
| Text Atlas | ≥ 18 | ≥ 3 | ≥ 60 | 4 |
| Demand Forecast | ≥ 15 | ≥ 5 | ≥ 98 | 3 |
| Free-threading lab | ≥ 18 | ≥ 19 | ≥ 379 | 7 |
| MILP Energy Lab | ≥ 35 | ≥ 15 | ≥ 525 | 6 |
Counts include recorded calls to each build model, including failed calls where logged. ≥ marks a minimum reconstructed from replies or model usage records; earlier network failures may be missing. Provider-access probes, validation runs without a model call and supervising-agent calls are excluded. A completed response can still contain incorrect code. Counts describe these builds, not model quality.
Historical Qwen RTX experiment · September 2026
Local generation used supplied AGILAB code, helpers and independent checks, with online Codex supervision. This historical experiment is distinct from the Qwen Flash editions; its timings are not a controlled benchmark.