Each agent is a self-contained application: its own entities, its own storage, its own routes, its own screen. They share a harness and a model catalogue — and nothing else. That is what lets one ship to production while the next is still finding its feet.
Code Live
An autonomous coding agent with a workspace of its own.
Give it a repository and a task. It explores, plans, edits, runs the tests, reads the failures, fixes them, and commits — in a three-panel IDE where you can watch every step and take the wheel whenever you want.
Three-panel workspace: explorer, editor, agent
Plan mode, permission rules and blocking hooks
Build, lint and test gates before work is called done
Checkpoints and one-click rewind on every step
Sub-agents, worktrees, task queues and skills
Live app preview and a browser it can drive
Chat Live
A general-purpose assistant that does more than talk.
Research, writing, analysis and everyday questions — with live web search, its own documents, and charts drawn straight into the conversation. Attach files and it reads them; speak and it listens without your voice ever leaving the browser.
Live web search with source-targeted queries
Generates PDF, DOCX, XLSX, PPTX, CSV and TXT files
Draws bar, line, area, pie and donut charts inline
Reads the documents you attach to a conversation
Vision input on models that support it
Projects, so long-running work stays grouped
Constructions Beta
A CAD/BIM assistant that draws while you describe.
Describe a building and watch it appear in 3D — walls, openings, levels and structure — with European and Romanian normative guidance applied as you go, and validation that catches what does not stand up.
Live 3D model built from plain description
EU / RO normative guidance in the loop
Structural and code validation with auto-fix
Runs as its own isolated pod
On the drawing board
The pattern is deliberately repeatable: a new agent is a new isolated domain, not a new prompt. These are the ones we are exploring next.