LEVEL 01 — AGENT INFRASTRUCTURE
Saguaro Agent
A vendor-neutral workflow, memory, and knowledge engine for AI coding agents — so the context built up in one session is still there in the next, across Claude Code, Codex, and Gemini CLI.
● PUBLIC SOURCE LIVEOpen sourceMCPAgent infraMIT
▸ BRIEFING
AI coding agents are sharp inside a single session and forgetful at every boundary that matters — a new chat, a different harness, a fresh checkout. The context an agent builds up evaporates exactly when it would be most useful to keep.
Saguaro Agent exists to fix that. It gives an agent durable working memory, a knowledge store, and a repeatable workflow, exposed through a stable interface that Claude Code, Codex, and Gemini CLI can all speak to.
▸ WHAT I OWNED
- Designed the model that separates workflow, memory, and knowledge — and the boundaries between them.
- Built the three MCP servers — saguaro-workflow, saguaro-memory, saguaro-knowledge — and the plugin surface agents call into.
- Kept it provider-neutral: OpenAI-compatible embeddings and chat, so there is no vendor lock-in.
- Shipped it as MIT open source with documentation and worked examples.
▸ THE HARD PARTS
- Memory that helps without poisoning. The "1% rule" — consult memory and knowledge whenever there is even a 1% chance it is relevant — keeps recall useful instead of noisy.
- One design, many harnesses. Supporting different agent environments without collapsing to a lowest-common-denominator interface.
- Control you can see. Keeping the agent autonomous while a human can still inspect, correct, and trust what it is doing.
▸ ARCHITECTURE
AGENT /
HARNESS
Claude Code, Codex, Gemini CLI
SAGUARO CORE
Workflow · Memory · Knowledge
REPO /
TOOLS
Code, files, project state
// The MCP layer is the control boundary — a stable, inspectable seam between the agent and the system.
▸ DECISIONS THAT MATTERED
MCP as the control boundary
Why: a stable, inspectable seam between agent and system beats bespoke glue rewritten for every harness.
Separate memory from knowledge
Why: short-lived working state and durable facts have different lifecycles and trust levels — collapsing them made recall worse.
Provider-neutral by default
Why: OpenAI-compatible interfaces keep the engine portable instead of welded to one model vendor.
MIT and public from day one
Why: the value is in being adopted and inspected, not hoarded behind a wall.
▸ WHAT SHIPPED
- A public MIT repository with the workflow, memory, and knowledge subsystems.
- Three MCP servers that Claude Code, Codex, and Gemini CLI can plug into directly.
- Provider-neutral embeddings and chat, plus the "1% rule" operating model and docs.
NEXT LEVEL
02 →11 Mile Labs Skills