Notes on agents, software, and work.
Systems / August 16, 2026 / 9 min read
AI playbooks: how to evaluate the workflow behind the label
AI playbook is an informal label, and recurring AI workflows can be assessed through scope, approvals, completion measures, and recovery.
Systems / August 16, 2026 / 10 min read
AI audit trail: how to judge whether an agent action is reconstructable
A useful AI audit trail connects the request to the acting agent, authority, execution, approval, verification, and persisted outcome.
Systems / August 16, 2026 / 10 min read
RPA vs agentic AI: choose by control boundary, not feature count
RPA fits predictable rules; agentic AI fits contextual choices, but only with explicit permissions, observability, state, and recovery.
Systems / August 16, 2026 / 10 min read
RAG vs agentic AI: what changes from retrieval to action
RAG retrieves evidence, agentic RAG changes how retrieval runs, and agentic AI acts across tools. Each step adds cost and control work.
Systems / August 16, 2026 / 11 min read
How does agentic AI differ from traditional automation?
Traditional automation follows a predefined path; agentic AI selects actions at runtime, so the right choice depends on path uncertainty and action consequence.
Operator notes / July 22, 2026 / 9 min read
Antigravity vs Cursor: choose on control and recovery
Leave with a decision you can defend: which tool suits your kind of work, what you can control before the agent writes to your repo, what you can recover when a long run fails, and how the launch-week verdict has aged, so you choose on cont
Operator notes / July 22, 2026 / 9 min read
Agentic AI programs: pick one by cost, depth, and skill
By the end you can place every serious agentic-AI program on a cost-and-depth ladder, judge each by whether it teaches the engineering that determines agent performance, and pick the right one for your budget, timeline, and goal, or decide
Operator notes / July 22, 2026 / 7 min read
Everything Claude Code: which repo is real, what to install
You leave able to confirm which everything-claude-code repo you're actually on, count its components against a named release tag instead of trusting a headline number, and choose a defensible install profile instead of running everything.
Operator notes / July 22, 2026 / 9 min read
Home Assistant MCP server: which one do you actually mean?
By the end you know which of the two things called a Home Assistant MCP server you actually want, how to connect it to your LLM client, and the one write-surface decision to make before you do.