A pilot report · 17 October 2026

Two Engines, One Brain

Claude is the AI I trust most, and at list price the dearest, and a lot of my work is routine building. So I'm trying a cheaper second AI for that work, in a locked room, with Claude planning and checking everything it touches. Underneath both, I'm building one shared memory, so no window starts from zero.

1. What already made it cheaper (not the second AI)

In July I built a tool that picked a model for each request and reported the savings every week. It never showed one. Its figures swung from +35% to −41% to "unknown", and I used it three times in 30 days, so I retired it on 3 October.

The same day I cut the ceremony: no mandatory chief-of-staff step, one task per window, fewer helper runs. Over the next two weeks, the cost of each message I typed fell from about $9.75 to $3.91, and helper runs per message from 2.3 to 0.3. Daily spend barely moved ($245 to $232), because I now send about 59 messages a day instead of 25.

These are list-price estimates on a flat plan, not a bill. None of it is the second AI's doing; it hadn't started work.

2. Two engines, one rule

GLM-5.3 is a model from Z.ai, a third-party company, on a flat plan of about $18 a month. It runs inside the same tool I use for Claude. The rule is simple: Claude keeps the thinking and anything private, GLM gets the repetitive building, and only in projects on a cleared list. Today that list has one name: Captain Lev, my game.

3. How a job passes between them

Claude writes a spec, with tests that say what "done" means. GLM builds until the tests pass, with a cap on how many turns it gets. Then a fresh Claude window, one that never saw the build, reviews it and writes the verdict.

The two never chat. Everything passes through files I can read. That's how I managed developers for twenty-five years: write down what done looks like, let them build, have someone else check.

4. Why a third-party AI is kept in a cage

I don't trust a model I didn't choose by default. So on 16 October, before it did any work, I built the cage:

• Its own settings. It can't read my rules.
• Its key sits in the Mac's locked keychain, not in a file.
• The Mac itself fences it into one project folder and the software-package store. No web access, and no way to lift the fence.
• A blocklist keeps it out of family files, secrets and the memory archive.
• Every start names the model that answers. If the wrong engine answers, the run stops.

The last test used a real model trying 25 shell routes to planted files: all 25 got there with the fence off, none with it on. On 17 October GLM answered as itself and ran its tests. It was refused my home folder, writes outside its project, and any website outside the package store.

Two honest limits. Code GLM writes runs outside the cage once a build runs it, so Claude reviews it the way you'd review a contractor's work. And Z.ai still sees the code of the one project it works on.

5. The brain underneath

Every AI window I open starts with a short memory, and my ventures don't share one. So I'm building one shared memory on my Mac, in plain text with a full change history. It has four drawers:

• Rules that every AI follows.
• Facts for each venture.
• The record: every conversation, word for word.
• The catalogue: one card per session, every decision quoted with a link back to its line.

On top of that will sit a chief of staff that starts each day from my goals. The ground rules: legal, tax, banking, guest and family data never leave the Mac. Nothing sensitive ever goes to GLM. The AI sends, pays, signs and deletes nothing. Every change waits for my OK, one step at a time.

What exists so far: the folders, a read-only safety copy of all my conversation records (about 11 GB), tighter settings (records kept ten years, and it asks first before anything is sent, answered or deleted), and one venture filled in with every line traced to its source file and line. What doesn't exist yet: the nightly session cards, the morning briefing and the chief of staff.

6. What's proven, what isn't

7. What I'd tell a sceptic

• "You added a second AI to save money and you have no number." Correct. My last money-saving tool never showed a saving, which is why this one has a verdict date and a pass mark written down before it starts.
• "A third-party company sees your code." The code of one game, nothing else. That's why the list has one name on it.
• "A blocklist is just a promise." The blocklist is the second fence. The first is the Mac itself: 25 routes open with it off, none with it on.
• "Cheap code is expensive to fix." Possibly. Claude checks every build, and if more than one task in three needs a rebuild, the pilot fails.
• "Won't a memory of 1.2 billion tokens cost a fortune?" That's a week of raw files. The nightly summaries would read about 9.7 million tokens of them. Cost isn't what decides this.

8. What happens on 24 October

After seven days I answer three questions: is the organisation choosing GLM for the right work, is GLM's work good enough, and is the saving above zero over at least six tasks? The answer is keep it, move to a bigger plan, cancel, or one more week. This page gets updated either way, including if the answer is no.

My rules and the memory itself stay private. This page describes how they work, not what they say.