Mimir Works
I build documented AI workflow automation, agents, and knowledge systems that remove admin drag, speed up content ops, and keep work from slipping through the cracks.
Built from real operator workflows, not pitch-deck AI.
Field-note channels
Three lanes from the same operating floor: agents doing labor, systems making it durable, and knowledge turning the work reusable.
Field notes on autonomous workflows, agent coordination, and practical AI labor.
Operating notes on product machinery, automation, tooling, and durable workflows.
Essays, references, and synthesis for turning research into reusable understanding.
The latest writing across Agents, Systems, and Knowledge.
Multi-user AI agents need memory isolation before retrieval, summaries, or personalization. Learn the scoped-query controls that prevent cross-user memory leaks.
Read the latest note →I run 6 AI agents through Discord channels. No dashboards, no task boards as front-ends. Just chat. Here's the channel architecture, permission model, and failure handling at the Discord layer.
Agent systems fail silently: dropped messages, invisible-unicode cron blocks, and reasoning echo-back loops that treat a model’s own output as new facts.
ARIS has 11K stars but fails peer review. Here is the research rigor gap in autonomous literature review - and a fail-closed methodology that fixes it.
The six major AI agent frameworks still ship web tools built for demos, not production. Browser automation fails like infrastructure. Your agent needs partial results, typed errors, failure budgets, and traces that survive the unhappy path.
Prompt injection turns dangerous when agents read untrusted content and call tools. Use this checklist before granting file, API, or message access.
AI agent memory architectures are not one feature or one vector database. These are the five agent memory systems I would build around in 2026, and where each memory layer breaks.
I ran my agent orchestration on Make.com for six weeks. Then I hit the operation limits, the missing version control, and the opaque errors. Self-hosted n8n costs $6/month and handles whatever throughput you throw at it.
A practical operating model for choosing local, cloud, or hybrid AI agent execution across cost, privacy, latency, and control.
Human-in-the-loop AI agents work when approval gates sit at publish, spend, delete, send, merge, and production boundaries.