Agent-Knowledge-Standard/AKS-Specification: Compiled domain knowledge, portable across agents and tools.

Compiled domain knowledge, portable across agents and tools. - Agent-Knowledge-Standard/AKS-Specification

Summary

The Agent Knowledge Standard (AKS) is an open specification for representing compiled domain knowledge as a structured graph that any agent or tool can query through a standard interface. Instead of treating knowledge as text chunks for vector search, AKS defines entities with typed relationships, confidence scores, and source traceability, exposed over a set of HTTP endpoints. Version 0.1.0 ships as a JSON-based spec that implementers follow so domain understanding survives across pipelines, sessions, and tools.

What it is useful for

AKS is useful when you need an AI agent to understand a domain and you do not want to redo that work for every new tool, session, or team. The README contrasts a knowledge base, which stores documents and returns text that looks related, with a Knowledge Stack, which knows the entities, their typed relationships, which relationships are well-supported, and what is missing.

Concrete examples from the spec:

• An agent answering a refund question can traverse from a Refund entity through Refund Policy, Refund API, and Manual Invoicing Exception rather than pattern-matching text chunks.
• A billing domain can be contributed to by both a support lead (policies, escalations) and a developer (API specs, ADRs); the compiled graph holds those facets together.
• Connected sources like Slack or Jira can sync into the compiled layer or be fetched live at query time, distinguished by the source_type field on responses.
• An MCP integration path is defined where a server queries the live Stack rather than dumping a giant file into context.

How engineers can use it

The spec is the deliverable. Engineers use it by either implementing AKS-compliant endpoints or building tools that call existing ones.

Documented endpoints:

• GET /aks/v1/entities to list entities, filtered by type and min_confidence
• GET /aks/v1/traverse for breadth-first traversal from a named entity, max_depth 1–5, with optional relationship_types filter
• POST /aks/v1/context to assemble a subgraph for a natural language query, returning entities, relationships, a traversal_path, and a HIGH | MEDIUM | LOW confidence label
• GET /aks/v1/export to produce a portable JSON Bundle, with filters like verified_only or max_entities for context-window-sensitive consumers
• GET /aks/v1/pages/{entity_name} for on-demand Markdown generated from the current graph
• POST /aks/v1/backfeed/flag and /approve to feed agent-discovered knowledge back into the Stack after human review

Every endpoint requires either a stack_id UUID or domain slug; missing both returns HTTP 422. A documented limitation is bundle scale: at roughly 20,000 entities and 40,000 relationships a full export approaches 50MB and becomes slow to parse, with a practical ceiling around 500 entities for project knowledge-base consumption paths. There is no install command; the repository ships no runnable code (the language field is empty), and the license is listed as NOASSERTION.