Models
The Extella map describes the platform Qwen model, connections using your own model keys, and local models on a device. The choice of model is separate from the tools.
Section 01 / AI
AI agent
Finds and calls programs for your task.
The agent connects a task to tools: it finds and runs saved programs, and helps create a new one when the needed tool is missing.
Why it helps
Useful actions can be saved and used again. Context, knowledge and rules help the agent work within the task you set.
What it contains
Memory and knowledge
Concepts store knowledge the agent needs. Save the lasting context of a task instead of explaining it from scratch every time.
Rules
Rules set boundaries for the agent. They record requirements for using tools and producing results.
Agent teams
Default and custom agents can have different tasks and sets of Experts. Profiles help organise this work.
What was measured
Extella measurements, September and October 2026. Each number holds under the conditions given.
- A saved expert pays for itself from the second run.
Measurement of 27.09.2026 on the qwen3.5:9b model: a skill spent 329 tokens on every run and solved 28 tasks out of 100. An expert spent 403 tokens once, when it was created, and then solved 100 tasks out of 100 without calling the model.
- The agent reuses what exists.
The agent searches experts by the meaning of the request, not by exact name. On tasks longer than a minute, taking a ready expert was 2.4 to 10 times faster than writing a new one. A ready expert repeats its own mistakes too, so its result is worth checking.
- With a short task language, a small model solves more tasks.
On the same set of 18 tasks, a small model solved 10 when it wrote the answer in the task language with checks, and 2 without checks. It used 5.7 times fewer tokens than Python with the same checks. For a large model the task language added almost nothing.
- The platform gives the agent its rules, memory and tools.
An agent on the Extella platform model with tools enabled completed 12 tasks out of 12 and 40 out of 40. An agent created through the API answers with a refusal until a provider and a model are chosen in the Extella interface.
Articles and demos
DemoSite with a brand bookA model writes a site block from your words; the handler on a device checks it against the brand book.What it shows: The model writes, the handler holds the brand rules: a block that breaks the brand book never reaches the page, and one site law re-prices every block at once.DemoLive siteA coffee shop site that arranges itself for each visitor by a short program a model writes from one sentence.What it shows: Personalisation you can check before it ships: every change runs against all 72 kinds of visitors in 0.3 ms, and you see whose page will change.DocumentationYour first programConnect a device, save a small calculation and run it with different parameters.
Related articles and examples.