Patrick Hofmann @patrickh0fmann

delta-mind.at Graz, Austria Joined March 2009

Tweets

37
Followers

13
Following

31
Likes

12

Patrick Hofmann @patrickh0fmann

6 months ago

Maybe a smarter agent would've solved this itself. Until then, I do it on the server. delta-mind.at/en/blog/from-p… #BuildingInPublic

0 0 0 6 0

View Details

Patrick Hofmann @patrickh0fmann

6 months ago

So I pulled widening out of the agent. The patterns now live server-side, defined by the user, before the agent requests anything. The client-side code isn't deleted. It sits in a branch, waiting.

1 0 0 5 0

View Details

Patrick Hofmann @patrickh0fmann

6 months ago

My agent doesn't like wildcard commands. When I built apes, I made sure he could request a broader scope — approve once, a class of commands free for an hour. Just because he can doesn't mean he does.

1 0 0 13 0

View Details

Patrick Hofmann @patrickh0fmann

6 months ago

The hard boundary stays hard. Only the crossing is audited. sudo gives you a season pass — configured once, valid without question. escapes gives you a single ticket — signed, audited, one-time use.

0 0 0 12 0

View Details

Patrick Hofmann @patrickh0fmann

6 months ago

apes run --as root -- whoami A grant is created. I see the command, the host, the agent on my phone. I approve or deny. The binary verifies issuer, signature, approver, host, command hash, and replay protection before anything runs.

1 0 0 11 0

View Details

Patrick Hofmann @patrickh0fmann

6 months ago

Today my agent is going to do something it isn't allowed to. It runs as its own OS user on my Mac Mini. No sudo entry. No password. Nothing in /etc/sudoers. That's by design.

1 0 0 29 0

View Details

Patrick Hofmann @patrickh0fmann

6 months ago

Same question, asked twice. Two different answers. Two releases of @openape/apes. The best debugging I know is: ask it why. Full story: delta-mind.at/en/blog/when-y…

0 0 0 5 0

View Details

Patrick Hofmann @patrickh0fmann

6 months ago

The combined lesson: tools that talk to AI agents need three things. 1. The content itself. 2. A structural metadata anchor (exit code, status flag) so the agent reads carefully. 3. An action shape that fits the agent's native primitives. All three. Each alone is not enough.

1 0 0 12 0

View Details

Patrick Hofmann @patrickh0fmann

6 months ago

When your agent doesn't do what you told it to, ask it why. Yesterday my AI agent ran a command, read the output, and told me "please approve in the browser." Then silence. The output had explicit instructions for what to do next.