Solentrex · July to August 2026
Two AI tools, one brain, and a written cost policy
Two vendors' coding agents, one rulebook and one memory between them, and a spending policy written the day a run got expensive.
What it was
I run two vendors' coding agents, Claude Code and OpenAI Codex. Each reads its own instruction file and keeps its own memory. Left alone, that gives you two versions of the rules that slowly disagree, work done in one tool that the other never learns about, and a memory that belongs to whichever vendor you are using this month.
Cost was the other open question. An agent can spread work across many parallel workers, and nothing stops that from getting expensive.
What I did
There is one canonical instruction file, AGENTS.md, the filename several tools already look for. The Claude Code filename is a symlink to it, so it is the same file and not a copy. The rules live in one folder that the file points at rather than repeats. Both tools read the shared memory at session start and write findings back into it, and a meaningful change gets a line in the daily log so the next tool picks it up. Wiring in a new tool later is one link.
I verified the round trip on July 5, 2026 instead of assuming it. Codex read the required start files, checked that my local environment was isolated, wrote its answer into the shared record, and appended its own daily log entry. The other tool read that entry in its next session.
The policy came out of a bad day. A research run fanned out to 105 agents that inherited the main session's model and used 3.4 million tokens, which spent the session limit in one shot. The rules are written down.
| Rule | Why |
|---|---|
| The main session's model is never inherited by a fan out | Inheritance is how one search becomes 105 agents on the most expensive option. Every agent is given its model explicitly. |
| Model per seat, effort per lane | Judgment work gets the strong model. Research runs at high effort and implementation lower, because implementation agents over engineer at high effort. |
| Never the smallest model | It is cheap and it invents things. |
| A standing adversarial verifier | One seat exists to contradict the managing agent. In a seven worker, fourteen hour run it refuted four of the manager's own claims. |
Getting non-engineers into the tools
The same principle works for people who do not write code. Give them one written rule that decides the hard case. The roof labeling that trains our geometry model is done by the people with the product context to judge the edge cases, not by an engineer. They work from a written guide with one rule at the top: a saved correction has to be a real correction. Open the model's guess, save it unchanged, and all you have taught it is that it was right.
The website copy worked the same way. The people who own the message wrote the words, and the pipeline treated their file as the source of truth.
What it produced
The memory is not locked to a vendor. Both tools work from the same rules and leave findings in the same place, and a third is one link away. The spending rules are a page anyone can read and hold me to, rather than a habit in my head. And the people with the most context and the least code do real work in these tools without an engineer beside them.