Notes on agents, software, and work.
Systems / October 27, 2026 / 11 min read
Copilot Studio multi-agent orchestration: a control-first decision guide
Copilot Studio multi-agent orchestration needs narrow roles, bounded tools, tested handoffs, traceable execution, and a defined stop path.
Read the essay
Systems / October 26, 2026 / 11 min read
Agentic AI in healthcare: evaluate the workflow, authority, and controls
Agentic AI in healthcare can be evaluated by tracing each trigger, action, human checkpoint, audit record, and rollback path.
Systems / October 25, 2026 / 10 min read
How to evaluate agentic AI developers before you build or hire
Evaluate agentic AI developers by inspectable task boundaries, permissions, state, evaluation, recovery, observability, and cost.
Systems / October 24, 2026 / 10 min read
Meta agent harness: choose the right control layer
A meta agent harness may coordinate agents, unify different harnesses, or optimize harness code; choose by control boundary, risk, and cost.
Systems / October 23, 2026 / 9 min read
Harbour agent harness: evaluating the Harbor Framework
Harbor evaluates agents through sandboxed tasks; structured harnesses fit process-driven work, while lighter ones may fit open-ended work.
Systems / October 22, 2026 / 11 min read
Cursor multi-agent orchestration: a control-first decision guide
Cursor multi-agent orchestration becomes controllable when tasks, ownership, worktrees, review, and quality gates are explicit before parallel work starts.
Systems / October 21, 2026 / 10 min read
MCP servers market: how to judge marketplaces before you install
MCP marketplaces are discovery layers, and the useful choice depends on identity, install detail, permissions, compatibility, and removal.
Systems / October 20, 2026 / 10 min read
How to build multi-agent systems with Anthropic
Build an Anthropic multi-agent system only after a simpler design falls short, then map state, handoffs, observability, and recovery.
Systems / October 19, 2026 / 10 min read
Open AI Agent SDK MCP integration checklist
An OpenAI Agents SDK MCP integration needs the right transport, scoped tools, fresh discovery, traceable calls, and explicit cleanup.
Systems / October 18, 2026 / 11 min read
Harness engineering for GitHub Copilot: audit the workflow, not just the model
GitHub Copilot's harness is judged by repository context, execution, verification, traceability, and human review, not model choice alone.