Every instruction your AI applications run on — prompts, skills, tools, harnesses, knowledge — versioned like code, approved like policy, scoped by department. Authored in English by the teams who own the work. Portable across any model, orchestrator, or runtime.
Harnesses in a repo, agent-config MDs in one runtime, prompts in another, help & SOPs in a wiki, guardrails in a security console, tools behind a gateway, enterprise data everywhere. Agents multiply across departments: duplicated, conflicting, untraceable.
When an agent does the wrong thing, no one can say which instruction caused it, who approved it, or how to roll it back. When code sprawled like this, we built Git and GitHub.
Every agent, prompt, skill, tool, and harness — across every runtime.
Inventoried, owned, and deduplicated. Two departments automating the same process differently becomes a merge proposal, not an accident. Sprawl becomes an org chart.
COO — sprawl under controlEngineering, HR, Finance — each governs its own base.
Propose, review, approve, publish, roll back — in plain English, without an engineering ticket. The controls finally reach the people who have the knowledge.
Department heads — no more waiting on the queueEvery action traces to a version. Every change is reversible.
An immutable trail of who changed what, when, and why — across every runtime. EU AI Act logging and SOX documentation become a report, not a six-week project.
CIO / Compliance — audit-ready by defaultVendor-neutral, standards-native.
Instructions serialize to SKILL.md, AGENTS.md, and MCP — plug into any model, orchestrator, or runtime. Mix Copilot Studio with Claude and n8n; switch without rebuilding what your company taught its AI.
Platform eng — no lock-in, git-syncedEnglish is the new code. Scopes are how it compiles.
Every instruction lives at one of five levels. A change lands at the narrowest scope that fixes the problem — and inheritance carries it exactly as far as it should go, and no further. Company-level skills, knowledge articles, and tools are shared downward: to every department, or only to the ones company stakeholders allow.
Each department lives in one review queue.
Suggestions arrive from AI analysis of the logs in your observability platform, from your own team, and from policy changes — each one a scoped diff with evidence. Approve, edit, re-scope, or reject. Nothing merges without an owner.
Engineers consume the result through one API. They keep git; you keep the wheel.
From AI · observability logs
Expense agent failed 12 runs on receipts over $500 — proposed update to Finance › Approvals F-3.
From policy · HR update
Parental-leave policy changed — 6 instructions across HR & Payroll need the new terms.
From engineer · Dev K.
Deploy skill contradicts release-freeze rule (company scope) — merge suggested.
the governed loop · scattered instructions → one base → every department → observability logs back
Knowledge, rules, tone, skills, and tools — drafted with a co-author that is itself governed by your base.
Every runtime consumes the base via API and MCP — per agent, per department, per scope.
Logs from the observability platform you already run, and every policy update, are checked against your instructions in bulk.
The AI suggests scoped changes to instructions, skills, and harnesses — reviewed, tested, reversible.
Suggestions come from your AI, your people, and your policies — but nothing merges without an owner.
Codex, Claude Code, Agentforce — one governed base behind all of them. Departments steer from the dashboard; you consume everything through three surfaces:
Retrieval governed by the search rules each department sets — filters, intents, and scope-aware permissions enforced on every query.
Create, update, and submit articles, rules, skills, and tool definitions programmatically — every write lands in the review queue, never straight to prod.
Mount the base directly into Claude Code, Codex, Agentforce, LangGraph, or any MCP-compatible agent — the harness travels with the standard.
Versioned instructions, approval records, and action-to-version trace for every agent in the estate — the exact evidence AI-governance reviews, EU AI Act logging duties, and SOX documentation ask for, produced as a report instead of a project.
Design-partner program open for AI platform & CoE teams running multiple runtimes.