The thing that clicked today: in a long agent conversation, a stale planning file is not a one-time read, it is a tax on every turn after it, and the cheapest model in the world does not change that.
Every working session with an AI assistant on the fleet starts the same way: read the project's contract, read the live handoff, read the roadmap, then work. What I had not priced is that those tokens do not leave the context. Everything read at orientation is sent again with every subsequent request, so a bloated planning file compounds for the whole conversation, and it does so identically whether the model is the expensive one or the cheap one.
Built / shipped
Measuring where the cost was. The roadmap was the dominant per-turn cost: 167 lines, of which about 140 were closed work clusters and an old 55-line progress log. Material that had been true for weeks and that no turn needed.
Archiving rather than rewriting. Every closed cluster and the whole pre-July-20 progress log moved verbatim into an archive file (22,557 bytes), and the live roadmap was rewritten to 8,610 bytes: the locked decisions, the open backlog, and an eight-line pointer back to the archive. Nothing was lost or reworded. The rule going forward is to archive a cluster the moment it closes, not in batches later, because the batch is exactly the thing you keep re-reading while you wait to do it.
Tiered orientation, written down. Read the live handoff every conversation; read the big reference documents only when the handoff's stamps say they have changed or the section actually edits them; read the roadmap only for replanning or at close. Those became standing rules in the project doctrine, shared across projects, rather than habits.
A capability audit before trusting the cheap model with more. The same redesign routed more routine read-only work to the free local model, so before any of that landed, a hardened audit of what that model actually does when asked: eight planned probes, each one a real command whose output had to be echoed verbatim, with file reads forbidden as a substitute. Seven of eight executed as genuine tool calls with exact output. One, a line-count probe, was silently dropped: the model's own written plan listed it, no matching tool call appears in the transcript, and it did not report a failure as the instructions required. Ruling: the routing is allowed, and a supervisor counts real tool calls against the plan rather than trusting the model's own completion claim. That caveat went into the same doctrine line as the routing.
Problems & fixes
The next day, a handoff arrived from a different AI strategy seat, a complete and internally consistent document claiming it had promoted the local model to a new one in production: a selector change, a delivery fix, continuity files updated. The question was not whether the document was well written. It was whether the work was real.
Verification against something the author could not have shaped after the fact. The relay's own audit listed about fifteen real runs across the previous eight hours whose tokens matched the document exactly, and the relay's live metadata for the local seat showed the new model name, a field the document's author had no way to fabricate. Four separate notes, read directly, confirmed the claimed content in each seat's own voice, consistent with those run tokens. Conclusion: the promotion was real.
Two things the same document got wrong or left out. One continuity note said the next task was a ratification the handoff itself had superseded earlier that day, a staleness from reading an older version. And more seriously: a session note whose filename implied a design had been ratified described work that had in fact stalled. The other seat had hit an evidence-reliability problem with the local model (an omitted table, false claims that files were absent), correctly refused to accept the model's negative claims as ground truth, and left the real next step undone. The filename said done; the note said otherwise.
A governance gap, filed not ruled. A second strategy seat with independent authority to fire jobs and write continuity files is a multi-writer problem the existing rules had not anticipated. It became an open decision for me, not something to settle in passing.
Decisions
Archive closed work the moment it closes. The live roadmap holds decisions and open items, nothing else.
Orient in tiers: handoff every time, reference documents on stamp change, roadmap at replan or close.
Cheap-model routing is allowed with a supervisor counting real tool calls against the plan. A dropped probe that is not reported as a failure is why.
Another agent's handoff gets the same treatment as a worker's verdict: corroborate against live system metadata, not prose.
Learned
Per-turn cost is model-tier independent. The biggest saving in the whole redesign was moving text out of a file, not changing which model read it.
A silently dropped step is worse than a failed one. Seven of eight done and one quietly skipped, with no failure reported, is the pattern to design against.
A detailed handoff is still a self-report. What settled it was a metadata field the author could not have written.
A filename is not a status. The note that looked like a completed ratification recorded a stall.
Still open / next
The governance question about a second strategy seat's write authority is open and needs a ruling, not a workaround. And the orientation tiers are written down; whether every seat actually follows them is something the next few sessions will show.