Notes on agents, software, and work.
Systems / September 29, 2026 / 10 min read
Context engineering skill: how to inspect, test, and trust an agent skill
A context engineering skill is an inspectable contract for activation, context, authority, evaluation, retained state, and recovery.
Systems / September 28, 2026 / 9 min read
Agentic AI coding tool: choose by control, verification, and recovery
Choose an agentic AI coding tool by matching task scope, permissions, verification, recovery, and usage costs to the work you will delegate.
Systems / September 23, 2026 / 11 min read
Human-in-the-loop AI SDK: how to choose for durable approval
A production-ready human-in-the-loop AI SDK can pause, preserve state, resume, time out, and record who approved what.
Systems / September 22, 2026 / 10 min read
Open source AI agents on GitHub: inspect the harness before you choose
Compare open-source AI agents on GitHub with one scorecard for control, state, evaluation, security, recovery, and cost.
Systems / September 21, 2026 / 10 min read
Model Context Protocol vs API: when to use MCP, an API, or both
MCP adds AI-facing discovery over APIs; choose a direct API, MCP-over-API, or both by caller, latency, authority, and operations.
Systems / September 20, 2026 / 10 min read
How to use MCP servers in VS Code: scope, permissions, and recovery checks
Use MCP servers in VS Code by matching each coding task to its capabilities, authority, approval surface, and recovery path.
Systems / September 19, 2026 / 10 min read
Pi agent harness GitHub: which repository and package do you need?
Pi is a layered coding-agent ecosystem with separate surfaces for the CLI, runtime, model API, interfaces, packages, and session state.
Systems / September 18, 2026 / 10 min read
OpenAI Agent SDK human in the loop: pause and resume
OpenAI Agents SDK human-in-the-loop control pauses risky tool calls, preserves run state, and resumes the same run after approval or rejection.
Systems / September 17, 2026 / 11 min read
MCP human in the loop: where to pause, approve, review, or block
MCP human-in-the-loop controls should vary by action: low-risk reads can run, reversible writes can get review, and risky actions pause.