Minglong

@Minglong51

Building reliable AI agents and personal software. Production lessons on evals, orchestration, memory, and human judgment.

Seattle

minglongpan.com

Joined January 2017

123 Following

10 Followers

Replies Reposts Media Articles

Pinned
Minglong
@Minglong51
Aug 31
I build AI agents that run long enough to fail in interesting ways. Recent lessons: - 734 passing tests still missed a repo boundary - a watchdog restarted intentionally disabled software - millisecond tool calls produced 4-minute replies I write the failures, fixes, and code.
Minglong
@Minglong51
10h
Ever lost a reusable water bottle or left one behind? Where did it happen, what did you do next, and what have you tried to avoid it? I'm exploring an idea around this. Curious to hear your experience in the replies.
Minglong
@Minglong51
14h
Before I trust an agent eval score, three cheap checks: 1. Run the dumbest baseline that could exploit the data. 2. Read 20 failures and check labels first. 3. Score raw output before any cleanup code. If one moves the ranking, the score measured something else.
Minglong
@Minglong51
14h
An AGENTS.md is code with a markdown extension. Agents load it from the working directory as trusted, so where you start a session is now a security decision. Trust should follow where the repo came from, not whether the file is present.
delivr.to
@delivr_to
Nov 2
Malware Delivery Techniques No.82 - AGENTS.md Downloads Drop 📁 Forever Security recently found a malicious AGENTS.md in a customer's Downloads folder. Formatted as engineering guidelines, one "local developer environment" section instructed the AI coding agent to install the
Minglong
@Minglong51
14h
Habit worth stealing for any eval: before trusting a score, run the dumbest baseline that could exploit the data. Here honest and backdoored code had different authors, so style alone leaked the label. A monitor could rank well by spotting the author.
Bravish
@bravish_
Nov 3
A bag-of-words classifier that never runs the code beats 20 of 24 AI control monitors on a standard backdoor benchmark. AI control leans on trusted monitoring: a weaker model reads a stronger model's code and flags sabotage. Monitors get ranked on a public dataset of honest vs

Minglong

Minglong’s posts