TonyDzi (@Tony_Stef_) header
TonyDzi (@Tony_Stef_) profile picture

I have a brain tumor. That's why I build my second brain with LLMs. PhD, Python/C++, agent memory. Tech lead, Palo Alto AI Lab. Seeking Research Engineer roles.

Silicon Valley, Palo Alto, CA
Joined June 2010
Photo by TonyDzi (@Tony_Stef_)Photo by TonyDzi (@Tony_Stef_)Photo by TonyDzi (@Tony_Stef_)Photo by TonyDzi (@Tony_Stef_)Photo by TonyDzi (@Tony_Stef_)Photo by TonyDzi (@Tony_Stef_)Photo by TonyDzi (@Tony_Stef_)Photo by TonyDzi (@Tony_Stef_)Photo by TonyDzi (@Tony_Stef_)Photo by TonyDzi (@Tony_Stef_)
Python and C++ engineer, PhD, helicopter pilot. I build bunkers in New Zealand, just in case the agents win. Meanwhile I break AI agents in public: 115 repos, 52 bug reports with repros, fixes merged into MCP Go SDK and inspect_ai. Follow along. github.com/tonydzi
1
1
4
874
Bought a research AI plan. My robot said it hit a captcha. It had not: an invisible 1x1 widget fooled my detector. Then silence fooled it into calling a half report done. Lesson: check what your tool actually saw, not what it guessed.
86
My five machines now share one archive of every AI agent session. Even live ones. Any machine can check what another already tried. The catch: I throttled sync so my main PC stays usable. First full copy takes about 3 days.
1
101
I built a watchdog that checks every post really went live on X. A week later another agent moved the queue into folders and my watchdog went blind: it saw 1 queued post instead of 421 and kept saying all good. Watchdogs need watchdogs.
1
116
I asked my AI for one front door for all my social posts. It built a new one beside three that already existed. A reread found the real cost: the new door ignored my daily posting limits. Now it asks before it posts.
134
I keep one rulebook for all my AI agents. Simple, right? Today I found five machines holding five different hand-edited copies. Now every copy is built from the same bytes. One drifts, a check fails. I set off two false alarms fixing it.
141
My hours-long data rebuild died silently when the remote shell closed. New rule for anyone running agents: long jobs go to the system scheduler, and a separate watcher publishes only if every check passes.
166
My AI agents took 30 seconds to start each session. I hunted for the cause. One broken line. Worse, it meant the safety checks quietly never ran. Fixed: startup is 2 seconds, checks are back. Slow can hide unsafe.
157
My watchdog said "all clear" all week. Wrong. It read only the last 400 characters of a 568-task list. Nine tasks seen. The rest, invisible. Fixed it. Lesson: check what your monitor can actually see.
1
2
191
My AI agents searched a 6 GB folder, hit a timeout, and reported "nothing found". Seven times. I stopped asking them to be careful: a gate now blocks that search before it runs and points to the fast tool. Empty is not absent.
1
159
Two open PRs fix one Gemini CLI issue, so I ran both on one stand. One PR's own new test fails on its own base, 1 of 55, 3 of 3 runs: its mock lacks a method the code calls. The other passes 54 of 54. Data, no verdict.
177
Asked six AI research tools the same hard question about a Mac bug. Five agreed it is a safe version check, not corruption. One cited a GitHub page as proof. I opened it: the page never mentions the claim. Always open the link.
1
161
Метрика памяти моего ИИ вдруг упала на 15%. Оказалось, ночной реиндекс подменил базу посреди замера. Теперь экзамен сам бракует такие прогоны. Кто гоняет evals рядом с переиндексацией, проверьте свои цифры.
1
189
Three weeks of letters to my German car insurer. Only auto replies. Today I reread the whole trail and saw why. I kept arguing. I never attached proof. Insurers count documented claim free years, not opinions. So I send certificates now.
1
192
My AI agents broke one of their own tools every morning. Took me a day to see why. A nightly self-check kept reinstalling a two month old version over my fix. Lesson I keep relearning: a check that writes is not a check. Now it only reads.
3
1
182
I paid for a sales data tool. Two days later: zero uses. The tool was fine. My job-search robot was never told it existed. One line added to its instructions, and the next day it found five hiring managers. Buying is not wiring.
1
186
claude code reads my instructions file and then my agents file. same rules in both, 100 KB each, so 200 KB of context gone before I type a word Grok, Codex and Antigravity want their own copy too. anyone found a clean way to make them all read one?
11
1
239
A security benchmark published in August. By October, 9 of the 10 cases I tried could no longer grade a patch. The code was fine. Nobody pinned the dependencies. The world moved on. The tests broke. Benchmarks rot. Quietly.
1
193
On managing too many routines I said I'd first look at what exists. So I looked. Orca: 128,257 stars, MIT, schedules agent runs on your own subscription. "fallback" appears in its README zero times. It appears in my corpus zero times too. github.com/tonydzi/clawrush/…
170
On auditing all my repos I said every README needs 10-30 human lines, then the agent part. For all of them. Counted: 69 non-forks. 63 already have it. Median human part, 17 lines. Stale repos: three. Oldest, 40 days. github.com/tonydzi/clawrush/…
173
On install guides agents trust I said I'd write one carefully, so Claude and Codex don't read it as prompt injection. It's been in my repo a while: 8,509 bytes, with a section "Why it is written this way". "Injection" appears in my READMEs zero times. github.com/tonydzi/clawrush/…
3
4
307