Push one model through everything
Make one chat remember the brief, write the post, debug the publisher, hold the approval rules, and recover every old decision after each compact.
Guth architecture · one real handoff, end to end
Ask for today's Instagram post. Rebecca owns social. If the publisher breaks, she pauses the post, calls Stan, and resumes with the repair—while you stay in the same conversation.
Follow this requestA scroll story showing how bounded workers, durable records, configured authority, and inspectable evidence keep one agent focused and accountable.
00 — The request
Make one chat remember the brief, write the post, debug the publisher, hold the approval rules, and recover every old decision after each compact.
Rebecca keeps the social work. Stan keeps the engineering work. The point guard moves the result between them and preserves the trail.
01 — The point guard
You do not have to choose a model, window, or machine. The point guard knows who owns the work and gives that specialist only the task, files, and authority the job needs.
02 — A real specialist handoff
This happened inside Guth on July 23. Rebecca had the social assets staged when she found the publisher was treating image slides like video. She held the post and called Stan instead of guessing around the failure. The alert crossed from her Codex task into Stan's active Claude task while he was already working another lane.
Reproduce the failed image handoff before changing code.
Open the publisher, failure record, media server, and test suite.
Fix PNG handling without weakening the approval boundary.
23 of 23 checks pass. The post is still held for Tony.
Repair received. I'm back in the same carousel, with the draft and schedule intact. I'll restage Instagram only, then send the exact post to Tony for approval.
03 — Source-linked memory
Her social lane keeps the brief, assets, schedule, hold decision, Stan's repair receipt, and Tony's approval state in one source-linked record. The next visit starts from the relevant state—not a retelling of the whole chat.
04 — Challenge retrieval
The approved record that helps a worker answer can also reveal why a candidate should not pass. Guth examines its support, governing evidence, freshness, contradictions, and authority before the work advances.
Retrieval supplies evidence. It is not truth, and it never grants permission. A verifier, reviewer, or configured gate decides the disposition.
05 — Change control
Sources, memory policy, tools, runtime, and test versions are bound to one manifest. Change any part and the digest changes, so the old and new setup cannot silently look the same.
Important: a matching digest shows that the recorded bytes match. It does not prove a claim is true.
06 — Permission gates
A protected action can stop at the Desk until an authorized person approves the exact action. The customer can make the gate stricter as the consequence rises, and the decision becomes part of the receipt.
The approval record can be a plain confirmation, a typed allowance naming the action and limit, or—when separately integrated—a signed authorization.
Illustrative policy mapping. The actual approval method is separately scoped, integrated, and tested.
07 — Receipts and status
Receipts record the exact work in order. They preserve the runtime, sources, decisions, permissions, and evidence state so the record can be inspected later.
A certificate commits to one scoped manifest and evidence state. The resolver reports whether that state is current, expired, superseded, or revoked.
The model is swappable.
The evidence system is the product.