OpenEchelon

# OpenEchelon

> OpenEchelon is an open-source, model-agnostic runtime for structured AI organizations. Instead of a flat multi-agent swarm, it models a hierarchy — CEO, executives, directors, managers, specialists, and workers — where each agent has defined authority, permissions, memory, reasoning budgets, and reporting lines. The human owner talks to the CEO and retains final authority. Agent identity is kept separate from model execution, so a single logical employee can be run by a local model, a frontier API, or a CLI coding agent depending on the task.

OpenEchelon is licensed under Apache-2.0 and is currently in the architecture and prototyping stage (pre-alpha). The Phase 0 organizational data model and a Resource Governor spike are implemented in Python and tested; there is no organization runtime and no released package.

## Definition

OpenEchelon: a model-agnostic runtime for structured AI organizations where persistent agents operate within hierarchy, authority, resource, memory, and information boundaries while reporting ultimately to a human owner.

## What makes it different

- **Hierarchy is a first-class primitive**, not an emergent behaviour of prompts. Reporting lines, departments, delegation, escalation, and review chains are modelled explicitly.
- **Agents are not models.** An employee is a persistent organizational identity with a role, manager, permissions, memory, and performance history. Execution resources are assigned per task by a Resource Governor.
- **Logical agents are separate from running processes.** A thousand registered employees does not mean a thousand running models.
- **Reasoning is a budgeted resource.** Actual reasoning is the minimum of what the task requires and what the role permits.
- **Cheapest sufficient intelligence.** Work starts on local or inexpensive models and escalates to frontier models only when justified.
- **Managers compress context.** Raw worker output does not propagate upward unchanged.
- **Provider independence.** Ollama, llama.cpp, OpenAI-compatible servers, Codex CLI, Claude Code, Gemini, and API providers sit behind adapters.
- **Security and auditability are architectural**, covering prompt injection, authority escalation, information boundaries, credential isolation, and approval gates for sensitive external actions.

## Key documents

- [README](https://github.com/NAME0x0/OpenEchelon/blob/main/README.md): project overview, organizational model, intelligence ladder, roadmap.
- [Immutable principles](https://github.com/NAME0x0/OpenEchelon/blob/main/docs/principles.md): 40 architectural constraints, non-goals, and the Architectural Test applied to design decisions.
- [FAQ](https://github.com/NAME0x0/OpenEchelon/blob/main/docs/faq.md): direct answers on what OpenEchelon is, how it compares to flat agent frameworks, and current status.
- [Contributing](https://github.com/NAME0x0/OpenEchelon/blob/main/CONTRIBUTING.md): what is useful to contribute during the architecture stage.
- [Security policy](https://github.com/NAME0x0/OpenEchelon/blob/main/SECURITY.md): threat model for autonomous agent runtimes and private reporting.
- [Phase 0 specification](https://github.com/NAME0x0/OpenEchelon/blob/main/docs/architecture/phase-0.md): the organizational data model, its acceptance criteria, and what is deliberately excluded.
- [Decision records](https://github.com/NAME0x0/OpenEchelon/tree/main/docs/adr): why Python, why architecture decisions are recorded, and why quality estimation is the component the cost thesis rests on.
- [Resource Governor spike](https://github.com/NAME0x0/OpenEchelon/tree/main/spikes/resource_governor): measured cost saving, false accepts, false escalates, and the limits of a structural sufficiency estimator.

## Status

Pre-alpha. Built and tested: the Phase 0 organizational data model (employees, departments, capability grants, tasks and delegation trees, typed messages, artifacts, reviews, execution resources, quota, routing decisions) and a Resource Governor spike. Modelled but not enforced: permission boundaries. Not started: the organization runtime, storage, and scheduler. No release has been tagged.

The Resource Governor spike measured an 85.7% cost reduction against a frontier-always baseline on labelled scenarios, with one false accept — a structurally complete but fabricated answer that a form-checking estimator cannot detect. That limit is documented rather than tuned away.

## Optional

- [Changelog](https://github.com/NAME0x0/OpenEchelon/blob/main/CHANGELOG.md)
- [Code of Conduct](https://github.com/NAME0x0/OpenEchelon/blob/main/CODE_OF_CONDUCT.md)
- [Citation metadata](https://github.com/NAME0x0/OpenEchelon/blob/main/CITATION.cff)