An interactive explainer on how LLM agents actually work — the agent loop, tool use, context, multi-agent systems, evaluation and failure modes. Fully simulated in the browser: step through an agent run and watch the transcript grow.
Portfolio & Writing
London · Cambridge
Max Khor
AI Engineer · LLM Agents · London
I build production LLM agents for coding and data workflows, currently at a global investment bank. My work is the machinery that makes agents reliable in the real world — execution loops, context construction, safe tool use, observability, and eval loops built from real-user failures. Before engineering, I read mathematics at Cambridge.
Projects
Published SDK for controlled data-agent workflows: Python execution only (no bash), handle-based state so large tables never hit the prompt, and a built-in eval harness with cost-tracked leaderboards. Deployed end-to-end as data-harness-ui — a chat workbench with GitHub OAuth, per-user budgets, and streamed chart rendering.
A benchmark harness for agents: traces, checkpoints, rollback, and a cost-aware leaderboard.
Writing
A practical guide to threads, locks, conditions, semaphores, blocking queues, graceful shutdown, worker pools, testing, and asyncio.
A personal agent that lives in WhatsApp's Message yourself chat: capture is explicit, triage is agent-assisted, the board is a markdown file in git, and the agent learns from corrections.
Technical Notes
First-principles explanations, worked implementations, and compact references for Python systems programming.