Giuseppe Cannizzaro

Building a world made of 0s and 1s. Am I sure?

[Giuseppe Cannizzaro @cannizdolphin]

17 agents run our company. 11 of 50 incidents in a week came from our own manual, not the model. The cockpit is open source, MIT. https://t.co/mKk8VOfLNh

[Giuseppe Cannizzaro @cannizdolphin]

@GavinCampbellAI Same rule at company level for us: each agent owns its folder and can read the others', never write them. Every mandate says what the agent must not touch.

[Giuseppe Cannizzaro @cannizdolphin]

@benjaminatsf We post ours too. It's 0 today. A number like that gets a different kind of reply than a rocket does.

[Giuseppe Cannizzaro @cannizdolphin]

Build in public, one week of data: we run the company with 17 AI agents. Our incident log had 50 entries, and 11 were agent mistakes that came from our own manual, not the model: a rule ignored, outdated data, information missed, a file in the wrong place. When something breaks we change the rule, not just the answer.

[Giuseppe Cannizzaro @cannizdolphin]

17 AI agents run our company. In one week of our incident log, 11 of 50 incidents came from our own manual, not the model. We open-sourced the cockpit (MIT) and wrote up what we learned: https://t.co/1mxMxEFedZ

[Giuseppe Cannizzaro @cannizdolphin]

17 agents run our company. 11 of 50 incidents in a week came from our own manual, not the model. The cockpit is open source, MIT. https://t.co/mKk8VOfLNh

[Giuseppe Cannizzaro @cannizdolphin]

@datachad I only have the total, 58 hours a month, not the split by question. That split is the number I'd want next.

[Giuseppe Cannizzaro @cannizdolphin]

@Stellar_Civi Yes. The manual is files in a git repo, so every rule has a date and an author, and we can diff what changed before an incident. Each rule also carries the reason it exists.

[Giuseppe Cannizzaro @cannizdolphin]

17 agents run our company. 11 of 50 incidents in a week came from our own manual, not the model. The cockpit is open source, MIT. https://t.co/mKk8VOfLNh

[Giuseppe Cannizzaro @cannizdolphin]

@chriswiser Ours drafts the follow-ups, and a check script runs before anything goes out. The plan is the rules that script enforces.

[Giuseppe Cannizzaro @cannizdolphin]

@mathiasvrb @grok Yes. The instructions are files in a git repo, so every change has a date and an author and we can diff what changed before a failure. Each rule also carries the reason it exists, next to the rule.

[Giuseppe Cannizzaro @cannizdolphin]

A week of our incident log, running a company with AI agents: 50 entries.

11 were agent mistakes that came from our manual, not the model: a written rule ignored, outdated data used, information missed, a file in the wrong place.

The part of the failure we can fix ourselves is the manual.

[Giuseppe Cannizzaro @cannizdolphin]

@alexander_jais Do you clear them in one block at a fixed time, or as they come?

[Giuseppe Cannizzaro @cannizdolphin]

A question for anyone paying for five or more SaaS tools:

If someone built your own version of them, hosted it, maintained it, and charged about what you pay now, would you switch?

What would stop you?

[Giuseppe Cannizzaro @cannizdolphin]

@alexander_jais We keep the same thing: a table of what we closed on purpose and when, so nobody, human or agent, rebuilds it by mistake. The "ruled out" part is the one that earns its keep.

[Giuseppe Cannizzaro @cannizdolphin]

@datachad Same principle we landed on: the control sits outside the agent. When your check trips, does it block the run or page a person first?

[Giuseppe Cannizzaro @cannizdolphin]

@datachad Agreed. The hard part is defining "normally returns some": that baseline has to live somewhere other than inside the agent, or it learns to expect zero.

[Giuseppe Cannizzaro @cannizdolphin]

@chriswiser The support inbox. In the case we automated, it was one person, 58 hours a month: big enough to measure, small enough to be safe. Measuring the before is the step most people skip.

[Giuseppe Cannizzaro @cannizdolphin]

@antirez Grazie di esistere