Direct answer · multi-agent projects
Give every model the same record, and make the record the memory. The hand-off is where multi-agent projects quietly fail: the incoming model starts blind, re-derives what the last one knew, and can contradict decisions you already made — without anyone noticing until it ships.
Measured on a real multi-file task, tiktoken o200k_base; the record is model-neutral. Method and caveats.
Because they have different strengths, different quotas, and different blind spots. Teams already do it: Claude Code for one kind of work, Codex or Cursor for another, a switch when a session dies or a rate limit resets, a second opinion when something smells wrong. The models aren't the problem. The hand-off is — each one starts blind unless something carries the context across.
Two things, one loud and one silent. The loud one is cost: the incoming model reconstructs the project by re-reading it — 22,897 tokens for the eight touched files on the measured task — before doing any new work. The silent one is worse: the incoming model contradicts decisions already made. It wasn't there when you chose the approach, the constraint, the naming — and nothing flags the divergence. It just builds, confidently, on a different version of history.
The measured reconstruction cost →Four habits get you most of the way:
These work. They also depend on every agent, every session, remembering to follow them — which is exactly the kind of discipline that erodes at 2am. The structural version doesn't erode.
Every agent works on one board: tasks are packets, work happens in governed lanes — a dispatch that opens the lane, signals as work lands, a return that binds to the exact dispatch that authorized it — all appended to a hash-chained record re-verified on every read. The incoming model reads that record instead of reconstructing the project: 1,013 tokens versus 22,897, measured. And because returns bind to dispatches, "who did what, under which instruction" stays a checkable fact rather than a recollection.
The same mechanism, for session death →That's VeriCommand Pro: a hosted drift review where a different vendor's model checks the work against the decisions recorded on the board. The review measured about 532 tokens versus a ~32,700-token median real agent turn — roughly 60× cheaper than one wasted turn. How many wrong-path turns it prevents is a modeled range, labeled as such, and savings only occur when a review catches drift. The principle is older than AI: nobody grades their own homework.
Measured inputs and modeled scenarios →The local core is free and model-neutral by design — the conductor between you and every AI build tool you use, so the next hand-off costs a record read, not a reconstruction.
