Build complex artifacts together.
A cross-domain implementation of Human-AI Co-Construction (HAI-Co²), where domain experts and LLM agents collaborate as equal partners to jointly shape objectives and solutions through natural-language interaction, shared structured artifacts, and preference-guided exploration of candidate solutions.
Both the human and the agent read and write the same set of structured artifacts.
Cooperation, not autocompletion.
Co-construction sits between full automation and pure assistance: the human and the agent meet as equal partners, with complementary strengths, and shape both the solution and the objective together (Dutta et al., 2025).
Equal partners, not assistants
The human is not a prompter and the AI is not a servant. Both sit at the table with complementary strengths: the human brings domain expertise, judgement and intent; the agent brings broad recall, fast exploration and tool use. Direction emerges from the dialogue, not from one side commanding the other.
Solutions through joint search
Problem solving is iterative search over a space of candidate solutions, with proposals, edits and counter-proposals flowing in both directions, rather than single-shot generation.
Objectives are co-constructed too
What counts as a good solution is not fixed up front. Preferences are elicited, refined and revised through interaction, instead of being assumed as a static reward signal.
Hierarchical, partial-by-design
Work moves freely between abstraction levels (intent, structure, detail) and partial drafts are first-class. Discoveries at one level can re-open decisions at another.
Natural language as the equaliser
Dialogue is the primary medium so that both partners can express anything, at any level of abstraction. Direct edits, rankings and categorical choices remain available as complementary channels.
Persistent, structured artifacts
Drafts, the plan and preferences live next to the conversation as durable, co-edited artifacts and remain available across sessions.
One architecture, several expert domains.
The same studio adapts to scholarly writing, scientific visualisation and cybersecurity triage: only the tools, schemas and validators change.
Related-Work Generation
Co-construct a survey section: candidate papers, a semantic graph of contributions, and a citation-grounded draft — guided by your preferences.
arXiv:2508.07955Chart Generation
From dataset and intent to a publication-quality figure: iterate on encoding, palette and annotations while the agent keeps the spec consistent.
CVE Retrieval & Triage
Co-construct a vulnerability report: retrieve candidate CVEs, rank by relevance to a target system and assemble an analyst-ready summary.
Artifact-centric. Domain-agnostic engine.
A small, domain-agnostic co-construction engine orchestrates the iteration loop, while each domain provides its own artifacts, refinement maps and validators. Observability, storage and the LLM client are reusable building blocks.
- Construction spacesPer-level artifacts (e.g. outline → draft → polished).
- Preference-guided searchCandidates ranked against decoded user preferences.
- Co-edited workspacePreferences, todos and structured artifacts, all read and written by both parties.
Open by design. Built to be extended.
HAI-Co² is released under the Apache 2.0 license and is developed in the open. The architecture is intentionally modular: the co-construction engine, the LLM client, the workspace tools and the domain-specific artifacts are all separable building blocks. You are warmly invited to study the code, adapt it to your own domain, file issues and contribute back.