MSTRMND
01—Insights

Notes from the operating layer.

Writing on agentic systems, runtimes, memory, infrastructure, and open source — from the people building them.
02—Latest
Governance · 4 min read

Untrusted content cannot authorize an action.

Web pages, emails, and documents can inform an agent. They cannot grant it permission. We fence external text as data so it never becomes authority.

Read article
03—All articles
Architecture

Design tools for the model, not for the API.

Wrapping every endpoint as a tool hands an agent an interface built for programmers. We design tools around intent, honest errors, and visible side effects instead.

5 min read
Infrastructure

Trace the decision, not just the call.

Request logs show what an agent did but not why, so we design agent observability around decisions: what the agent saw, what it chose, what it was allowed to do, and who approved it.

5 min read
Agentic systems

A handoff is a contract, not a transcript.

When one agent passes work to another, the context that matters is rarely the conversation. We design handoffs as explicit contracts that state the goal, the evidence, the limits, and who owns the result.

5 min read
Memory

A memory without a source is a rumor.

Agent memory fails quietly when facts lose track of where they came from. We attach provenance to every memory at write time, so recall can be weighed, corrected, and forgotten on purpose.

5 min read
Governance

Approval is an interface, not an interruption.

Human approval gates are where agent systems earn trust. We design them as a first-class surface — scoped, specific, and cheap to answer — instead of a pop-up bolted on at the end.

5 min read
Agentic systems

Models change. The intelligence layer persists.

Why we treat providers as interchangeable and put our engineering into context, memory, and governance instead.

8 min read
Architecture

What an agent harness actually needs to run in production

Tools, skills, evals, approvals, and durable state — the parts most agent demos skip, and how we build them.

7 min read
Memory

User-owned memory: graphs, ranking, and identity loading

The design behind mstrmnd-core’s memory layer, and why local-first Obsidian was the right reference adapter.

6 min read
Open source

One doctrine for every agent: the case for mstrmnd.md

How a single canonical knowledge base keeps Claude Code, Codex, Cursor, and Gemini CLI aligned on the same company.

5 min read
Infrastructure

Cross-platform by default: shipping to Linux, macOS, Windows, and mobile

Containers, scripting, and CI patterns that let one team ship the same system everywhere.

6 min read
Research

Visualizing the operating layer in WebGPU

Notes from building mstrmnd-layer with vgpu — turning “ambition to execution” into a rendered field.

4 min read
04—Let's talk

Ready to build what's next?

We're always open to new conversations. Let's create something exceptional together.

Start the conversation