Skip to main content
HAI-Co² (Human-AI Co-Construction) is a reference implementation of a framework in which a person and an AI agent co-create one shared solution, not a chatbot that answers, but an equal partner that edits a typed, branchable workspace alongside you. This site documents the system end to end: what the agent sees on every turn, how the shared workspace and typed artifacts are wired, how conversations branch like commits, and the architecture decisions behind it all.

​
Get started

What is HAI-Co²

The framework, the four shared surfaces, and how co-construction differs from RLHF and assistance games.

Local installation

Run the full stack on your machine with Docker Compose, then log in to the studio.

Demo walkthrough

See the studio exercised end to end: charts, preferences, a working document, and a plan.

​
How it works

Architecture

System topology, data stores, the per-thread tool model, and the decisions behind them.

Agent core logic

How the agent runs a turn: per-turn context, the ReAct loop, tools, the SSE stream, and artifacts.

Tools

The agent’s workspace tools, each documented with its arguments, returns, and the prompt it sees.

Artifacts

The typed-artifact pattern and the produce, persist, stream, render lifecycle of the centre panel.

Conversation branching

The turn-as-commit model: fork any past turn into a new thread; restore without destroying history.

Tracing & data collection

Phoenix tracing plus the replayable first-party dataset HAICO records for every co-construction.

​
Reference

Features

The end-to-end catalogue of the Co-Construction Studio’s capabilities.

Decision records

The architecture decision records: the why behind the structure.

API reference

Drive the whole loop over HTTP: auth, the endpoint groups, and the generated endpoint reference.

​
Try it

The live application runs at haico.gr. To run it yourself, start with Local installation; to see it exercised end to end, follow the demo walkthrough.