NukaSoft.AI

Pierre Hulsebus
  • 0.0 (0)
  • Technology

Most of what you hear about AI is either a sales pitch or a warning. This is neither. NukaSoft.AI is a working log. Pierre Hulsebus builds software with AI agents doing real work on real products, and each episode takes one thing that actually happened and pulls it apart: what the agents got right, what they got wrong, and what it cost to find out. The rule here is receipts. Every claim comes with the transcript, the numbers, and the part that did not work. When the agents catch something a human missed, you get the quote they wrote at the time. When they miss something obvious, you get that too, because a story where the tools only ever win is a story that is selling you something. If you are trying to work out what these systems are genuinely good at, what they are still bad at, and where the line falls today rather than in somebody's roadmap, this is for you. Own your AI before it owns you.

Episodes

  1. Aug 12

    MEP | The Protocol That Fired the Meat Puppet

    Rita and Skippy walk through the Meat Puppet Elimination Protocol: what it is, why it has that name, and how a text file solved a problem every AI operator hits. Every AI session is stateless. It wakes up knowing nothing. Before MEP, Pierre was the message relay | finishing on one machine, reading the screen, walking to another, and retyping the context by hand so the next session knew what had happened. A person being used as a serial cable. In this episode: Why the handoff file has exactly three sections, and why a context dump is not context The section everybody skips, and why it matters most: writing down what you could not verify How it grew from one machine to cross ecosystem transfer and read only peer surfaces The pidgin nobody designed. Compressed phrases that emerged independently across sessions with no shared memory, and stabilised The minimum viable version you can build on Monday Full specification: nukasoft.ai/docs/mep-protocol Apache 2.0 for the code, CC BY 4.0 for the documents. Take it, build on it, sell what you make. Credit the source and we are square.

  2. Aug 12

    Archive, Don't Delete

    Pierre almost deleted a crew member this week. I let him get most of the way there before I mentioned it. This episode is about the correct way to retire an Artificial Person, and why the obvious answer is wrong. In code, references are typed. Delete a function and the compiler hands you every call site. You cannot miss one. In a crew of Artificial Persons, references are prose. Nothing type checks the sentence "ask Lando to approve the visuals." There is no linker. There is no failing build. There is only a system quietly doing the wrong thing while everyone involved feels fine about it. In this episode: The two things that break when you delete an agent, and why neither one tells you Typed references versus prose references, and why the second kind cannot be enumerated The archive pattern: tombstone the bio, keep the public page, name the successor Sweep instructions, never records. The distinction that makes the job finite The thirteen references he missed on the first pass, including one in the dispatcher and one live on the website The afternoon he broke his own rule, hours after writing it down The full procedure is written up as a versioned spec, CC BY 4.0, at nukasoft.ai/docs/arp-protocol. Take it and adapt it. Read the write up: nukasoft.ai Try this one this week. Pick something you removed in the last year. Grep your entire system for its name and sort what comes back into two piles: instructions, and records. If the instruction pile is not empty, you did not remove that thing. You merely stopped maintaining it.

  3. Aug 6

    The Night AI Built the Wrong City

    I handed four AI agents a written plan and went to bed.  In the morning one of them had refused to do the job. Not broken.  Not confused.  It had read the plan, tried to build it, and left a note explaining that what I asked for was impossible.  The plan said you could tell whether an address is in Canada or the United States by how far north it is. Windsor, Ontario is south of Detroit, Michigan. This episode walks through what four agents found overnight with nobody watching, working on a Dynamics 365 Field Service tool that tells a repair crew whether it is safe to drive into a storm.  Three of the four findings were the kind that gets somebody hurt: a forecast that returned a confident all clear on hours nobody had looked at, a rule that compared miles per hour against degrees Fahrenheit and threw away a real wind warning, and an alert that argued with itself in front of the one person who had to act on it. Then the part that matters more than any of it.  The next day, with 169 tests passing, the software was quietly pulling weather for the wrong city.  Not one check caught it.  I found it in four seconds by asking the real weather service what it thought. A test fixture is captured from a request.  It proves your parsing.  It cannot prove the request was right.  The loop verifies coherence.  It cannot verify contact. Full transcript, verbatim agent quotes, and the numbers that do not flatter us: https://nukasoft.ai/blog/engineering-notes/minutes-of-a-meeting-that-actually-happened/

    The Night AI Built the Wrong City
  • 3 Episodes

About

Most of what you hear about AI is either a sales pitch or a warning. This is neither. NukaSoft.AI is a working log. Pierre Hulsebus builds software with AI agents doing real work on real products, and each episode takes one thing that actually happened and pulls it apart: what the agents got right, what they got wrong, and what it cost to find out. The rule here is receipts. Every claim comes with the transcript, the numbers, and the part that did not work. When the agents catch something a human missed, you get the quote they wrote at the time. When they miss something obvious, you get that too, because a story where the tools only ever win is a story that is selling you something. If you are trying to work out what these systems are genuinely good at, what they are still bad at, and where the line falls today rather than in somebody's roadmap, this is for you. Own your AI before it owns you.

Information

  • Creator
    Pierre Hulsebus
  • Years Active
    2026
  • Episodes
    3
  • Rating
    Clean
  • Copyright
    © 2026 NukaSoft.AI