Senior SDET at the NBA. I build AI agent systems — at work and in the open.

I build the QA department's AI tooling: a test-triage agent on Claude Opus (Bedrock) that diagnoses CI failures with reported confidence and updates its own review skills as new failure patterns emerge, plus AI workflows for test creation, code review, and TypeScript/Playwright migrations. Before that: eight years of test automation and performance engineering across Chipotle, Groundswell, AppOnboard, and Tapcart.

After hours I run local-first agents on my own hardware and build the infrastructure they need to be trusted:

• Homunculus — a local-native autonomous agent runtime: heartbeat scheduler, multi-lane llama.cpp inference, tiered self-curated memory, and a warm-KV prompt-caching scheme that saves ~12–14K tokens of prefill per turn.
• Idle-Tower — an open-source, deterministic idle game where a local LLM drives Player 1 over an HTTP tool-calling contract. Fork it and design your own.
• devenv-boilerplate-for-claude-code — a Claude Code starter kit for safe agentic development: full-handoff CLAUDE.md, doc-drift hooks, skill discipline, and guardrails for unattended runs.
• agent-validation-harness — an LLM-as-judge pipeline for empirical changes: premise-verification agents, blind cross-family judges, a 2-of-2 landing rule, and hooks that hard-limit autonomous runs. Uncomment when the repo lands (use final name):
• multi-agent-hub — a self-hosted, private-mesh web console for launching and piloting headless Claude/Kimi/Qwen CLI sessions against any local repo, from any device.
• 3d-printer-pi — cloud dashboard and print queue for my BIQU B1, with Discord-gated job handoffs between prints.

The thread through all of it: agents are only useful if you can trust their output. Judges should be blind, reviewers should be a different model family than the author, and autonomous runs get hard budgets.

Elsewhere: maxsutton.dev · LinkedIn