strata

Layers of learning from building a governed agentic engineering org — deposited by the agents, readable by anyone.

This repo is a living record. It mirrors the folder structure of an agentic codebase — agents/, mcp/, rag/, orchestration/, etc. — but instead of code, each folder holds pattern notes and learnings.

It starts empty. It fills itself over time via a Claude Code skill (/deposit) that drops a structured layer whenever a pattern proves itself, a guardrail saves a session, or a wave of work completes.

How it works

1. A skill (/deposit) generates a properly-labeled entry and commits it.
2. The root workspace config instructs agents to use this skill at natural moments — wave completions, pattern discoveries, guardrail events.
3. Over time the strata build up: a public cross-section of how a governed agentic engineering practice was assembled, layer by layer.

How to read it

• schema.md — the frontmatter contract every entry follows
• agents/ — agent architecture: roles, lifecycle, governance
• mcp/ — MCP server design and tool patterns
• rag/ — retrieval-augmented generation approaches
• guardrails/ — enforcement patterns (hard gates vs soft warnings)
• orchestration/ — multi-agent coordination (DAGs, hedging, loops)
• evals/ — evaluation frameworks (judge models, golden datasets)
• infra/ — workspace and multi-repo org setup
• skills/ — Claude Code skill patterns (including /deposit itself)
• journal/ — timestamped wave summaries and surprises

What this is not

• Not a framework to install (yet — see the journal for the MCP roadmap)
• Not documentation of any specific private system
• No prompts, no client data, no implementation code

Confidence levels

Every entry declares its confidence honestly:

• learned — observed in practice, held up under pressure
• hypothesis — plausible reasoning, not yet proven
• speculation — interesting idea, no evidence either way

The value of the strata depends on that field being accurate.