#Agents
Essays filed under this hashtag.
Loading essays.
A practical guide to choosing local stdio or remote Streamable HTTP MCP servers, with security and operations tradeoffs, failure modes, and production checklists.
A minimal AI stack for solo founders: one host, one coding assistant, four roles, narrow MCP or native tools, and a human approval gate for consequential actions.
Company facts change too often to hide inside model weights. Use memory, retrieval, access control, and approval gates to teach an agentic OS what your company knows without retraining a model.
How marketing agencies can run one AI OS per client without treating workspaces as security isolation. Covers retrieval boundaries, MCP authorization, scoped credentials, approval gates, and failure modes.
Build vs buy for AI agents is a question about who owns auth, memory, evals, retries and the pager. Here is the cost of each side, with the assumptions written down.
Four categories, placed on six things a buyer can check: what you install, who runs the model, what it remembers, what it may do alone, what it costs, how it fails.
A profile of the six places an agent system spends tokens, how to measure each one, and the fixes in order of payoff, with the vendor numbers cited.
Fine-tuning vs RAG is a diagnosis, not a preference. Fine-tuning moves behaviour and format. Retrieval moves facts. Here is how to tell which one you need.
AI governance, here, means a person approves before an agent publishes, sends, or writes to a customer record. This is not a legal opinion.
AI agents for marketing teams run inside Growth OS on the host you pay. The OS looks for the gap. It does not replace the person who approves the send.
An AI operating system is specialized agents, memory, and skills on the host you already pay. It is not a chatbot, and it is not software you install beside your tools.