Assemblix is the builder and runtime for AI training exercises. Methodologists edit prompts and branches on a canvas. A shared student profile carries progress from one drill to the next. Soft Skills Lab uses it for 2,000 AI negotiation sessions every month.
2,000 AI training sessions every month at Soft Skills Lab
Hard client, three rounds
Three bottlenecks keep methodologists waiting on engineering. Assemblix removes each one.
A single prompt edit sits in the release queue for a week. By the time it ships, your competitors have pushed ten iterations of theirs.
The canvas carries the agents and the state machine they share. A non-coder wires them together. A fresh scenario goes from empty to running in about ten minutes, and engineers stay on the product.
When a roleplay derails, nothing tells you why. Someone digs through logs, adds prints, restarts the session, and spends a week chasing a bug that lives in one line of a prompt.
The session view shows which branches the agent chose, which state updates fired, and which prompt actually ran. A methodologist finds the bad line in about two minutes and edits it in the builder.
Exercise three has no idea how the learner did in exercise one. Every session starts from scratch. Nobody sees the arc.
Pass `client_id` once. Every agent in the library reads the same profile. The qualifier's score from exercise one feeds straight into exercise two.
Soft Skills Lab is a negotiation school with 10,000 students and instructors teaching at VSE and Skolkovo. Until last year, every AI exercise went through engineering. Now the methodology team writes and ships them directly.
The old setup had twenty AI exercises that refused to scale. Five people burned about ten hours a week keeping them alive, and every new scenario meant days of developer work on top of that. The content queue never got any shorter.
The team moved every exercise onto Assemblix and handed methodologists direct authoring rights to the builder. Prompt changes now ship the same afternoon they are written. Engineers dropped out of the content loop entirely.
A new exercise now ships in ten to thirty minutes. The library grew from twenty exercises to more than fifty. Two thousand AI negotiation sessions run every month, and methodologist support time fell from ten hours a week to under one.
Numbers come from Soft Skills Lab's internal reporting. The full write-up covers methodology and before-after screenshots.
The methodology team designs the scenario. A developer wires up one endpoint. After that, every change lands through the builder and every session goes through the inspector.
Drag the nodes into place and connect them. Every node lives on one canvas: the agent, the branching condition, the tool call, the state update. The methodologist writes the prompts directly.
curl -X POST \
https://api.assemblix.ru/v1/exercise/run \
-H "Authorization: Bearer $KEY" \
-d '{
"message": "I'd like to discuss the contract terms.",
"chat_id": "sess_2xa9", // chat_id scopes this conversation
"client_id": "student_42" // client_id scopes the student profile
}'Your backend sends a POST with chat_id and client_id. chat_id scopes this conversation. client_id scopes the student across every exercise in the library.
Every session records which branch the agent took, which state updates fired, and which prompt actually ran. When a drill starts to drift, a methodologist opens the session, finds the bad line, and edits the prompt in the builder.
Everything a methodology team needs to run training like a product. Orchestration and observability ship inside the box.
One client_id is one student profile. Every agent in your library reads and writes to the same state. A qualifier score from exercise one changes the opening of exercise two automatically.
For example,The qualifier scores objection handling at 62, and the next drill opens on a harder buyer.
A drag-and-drop canvas for the full scenario. Methodologists edit prompts, branches and state updates without waiting on a release.
For example,A trainer tweaks an objection mid-cohort and the next student sees the new version.
The session view records each branch the agent chose, each state update that fired, and each prompt that actually ran. You can replay a sideways session and find the bad line in about two minutes.
For example,A drill starts drifting toward escalation, and the session shows the agent jumped to the wrong branch at step three.
OpenAI, Anthropic, Gemini, GigaChat and DeepSeek are all wired in. You can switch a whole scenario to a different model in about thirty seconds. Bring your own API key and Assemblix stops metering the traffic.
For example,Run the same drill on two models, compare the sessions side by side, keep the one that holds the rubric.
Two shapes we see over and over, and two we see often enough to mention.
AI scenarios for hard negotiations, crucial conversations and feedback drills. The trainee talks to the agent. The qualifier scores the round. Over several exercises the student's profile grows on its own.
Agents play the role of a difficult buyer. The shared profile tracks which objections a rep has already handled and which ones still need work. New hires practise alone, reviewers read the transcripts later.
A chain of agents runs the competency interview. Each agent focuses on one area. The shared profile builds a full skills picture of the candidate across every round.
Assemblix runs the agent embedded in your product. PMs edit the scenarios in the builder. Every conversation is inspectable in the session view. Iteration on a single prompt takes minutes.
Five questions every training team runs through before picking a stack.
| What you need | ChatGPT + Custom GPTs | Dify / n8n | Custom code | Assemblix |
|---|---|---|---|---|
| Student memory between sessions | manual | build it | Out of the box | |
| Shared memory across multiple agents | build it | Out of the box | ||
| Methodologist edits content without code | Out of the box | |||
| Every agent step visible in a session view | build it | Out of the box | ||
| From idea to production | hours | hours to days | weeks to months | minutes |
Credits apply only to the built-in Assemblix keys. Bring your own OpenAI, Anthropic or Gemini key and work without our limits, paying the provider directly.
Run your first exercise end to end.
Run your first training program for a cohort.
Roll training out across the whole team.
For teams with SLA and compliance requirements.