RinDig/icm-architect

What it solves

Multi-agent setups often rely on orchestration code to define steps, context, and state, which makes them hard to understand and debug. icm-architect replaces that code with a plain folder structure: numbered folders set the sequence, hierarchy limits what context an agent sees, and markdown files hold state. A person can open any folder and see what the system is doing, and one agent reading the right files at the right moment can do the work of a multi-agent framework.

How it works

It works in two modes. Build extracts the structure already implicit in how you describe your work (the stages, checks, stable vs. per-run parts) and creates the smallest workspace that carries that structure. Restructure audits an existing folder, classifies each file into types (catalog, contract, factory, product, dead), finds dependencies, proposes a migration map for approval, then migrates and validates.

Every workspace is built from three primitives: a sequence (numbered steps, each output feeding the next), a shelf (a folder of same-shaped files), and a catalog (a small file that points down). These combine into five forms: Pipeline, Record library, Knowledge bundle (including one built from transcripts with every quote checked against its source), Context map, and System map. The forms can nest.

Each step ends with a check (a person, a script, or a small model) so errors are caught before they amplify. The final validation is the walk test: an agent with no memory must orient, act, and report status from the files alone. The workspaces are plain folders and markdown with AGENTS.md as the entry file, and a one-line pointer is added for agents that expect a different filename.

Who it's for

People building agent-driven tools and workflows who want a simpler, more transparent alternative to orchestration code. It works in any agent that reads a SKILL.md file, including Codex, Gemini CLI, Hermes Agent, Claude Code, Cursor, and GitHub Copilot. It also helps teams restructure an existing repo or vault into an agent-friendly, interpretable layout.

Highlights

• Two modes: build a workspace from a description, or restructure an existing folder with auditing and migration.
• Built on three primitives and five reusable forms that nest, keeping small workspaces minimal.
• Each step ends with a check, and the finished workspace passes a “walk test” to ensure a fresh agent can navigate it.
• Vendor-neutral: plain folders and markdown with AGENTS.md, plus pointers for agents that expect other names.
• Optional Python scripts check links and input paths, draw workspace graphs, generate indexes, and verify quotes against sources.