# Myagi - extended context for LLMs

## Identity

Myagi is a long-lived intelligence substrate designed to compound capability across domains, tools, contexts, and time. The project root governs:
- Target architecture and module boundaries
- Integration points and dependency graph
- Build order and milestone verification
- Capability Fabric evolutionary learning subsystem

The wrapper does not modify child projects. Child projects (services and packages) are standalone modules.

## Current state

Phase 0.3 - Capability Fabric vertical slice proven (as of 2026-02-12). All 9 architectural invariants validated. Monorepo consolidation complete. No autonomous task has executed yet. The system is intentionally in construction phase until first governed goal is approved.

Shipped milestones:
- v0.1 Governance Foundation (2026-02-02): contract framework, validation suite (195 automated tests), fail-closed safety guarantees.
- v0.2 Operational Verification: planned next - prove runtime execution under supervision.

## Conceptual model

The conceptual hierarchy, from highest to lowest level of abstraction:

1. **Executive Function** - top-level controller for goal management, planning, task decomposition, resource allocation. Distributed across Intelligence Gateway, orchestration components, and Control Plane.

2. **Execution Contexts** - runtime instances applying specialties to goals with delegated authority. Ephemeral, stateless, revocable. Disposed without state loss.

3. **Specialties** - design-time capability bundles clustering skills, heuristics, evidence, and trust requirements into coherent institutional knowledge assets. Versioned, governed, evolved through institutional learning. Consumed by execution contexts, never modified by them.

4. **Skills** - composed capabilities combining tools with domain logic. Procedural memory in cognitive-architecture terms. Reusable across specialties.

5. **Tools** - raw, volatile primitives. Versioned and frequently changing. Never authoritative. Provided by runtime, MCP servers, external libraries.

## Capability Fabric

Memory-centric evolutionary subsystem responsible for turning repeated attempts into durable strategies through large-scale experimentation and deterministic evaluation.

Core invariant: deterministic judgment, non-deterministic exploration. Execution may vary; evaluation never does. Institutional memory over individual learning. Evolutionary ratcheting through cautious promotion.

Primitives:
- **Challenges**: atomic, verifiable problem units with explicit success criteria.
- **Strategies**: mutable, versioned instructions for addressing a class of challenges. Tracked with evidence (success rate, cost, latency, failure modes).
- **Operators**: ephemeral, stateless executors. Probes, not learners. Execute, report, dispose.
- **Execution Attempts**: immutable records of operator executions. Append-only.
- **Fitness Gradients**: continuous, time-weighted, context-aware confidence signals derived from attempts.

## Key principles

1. **Capability earned, not asserted**: specialties accumulate evidence; agents do not claim capability they have not demonstrated.
2. **Revocability**: every execution context can be terminated without state loss; no autonomous accumulation of authority.
3. **Supervisory control**: humans remain in the loop; first governed goal requires explicit approval.
4. **Institutional memory**: learning lives in specialties, not in disposable execution contexts.
5. **Construction-phase honesty**: nothing is offered until it has been verified end-to-end.

## What runs on myagi.bot

- Marketing site (this domain): public communication and contact.
- /open-agent-registry: separately deployed open registry scoring sites for agent-readiness against a transparent rubric.
- /api/mcp: Model Context Protocol endpoint exposing query_platform and submit_inquiry tools.
- Static well-known endpoints implementing the full agent-protocol surface area: agents.txt, agents.json, agent-card.json, OpenAPI, OAuth discovery, RSL.

## Public commitments

- All agent-access signals on myagi.bot are honest. We do not publish OAuth discovery unless OAuth works, OpenAPI unless paths resolve, MCP unless tools answer.
- The Open Agent Registry rubric is public, transparent, and adversarial: the registry will score myagi.bot itself the same way it scores every other site.
- All content under myagi.bot is licensed for AI training and inference subject to the terms in /.well-known/rsl.xml. Attribution required. No paywall yet.

## Contact

contact@myagi.bot. The contact form on the homepage and the MCP submit_inquiry tool both forward to the same inbox.