Every subject gets a page
17 September 2026 — we moved 'file a page for each thing a source is about' into one shared instruction every agent carries. Measured with our old per-agent note removed to isolate it, the shared instruction lifts page-worthy people filed from 6-in-10 to 10-in-10 — matching the note it replaces, now from one place, with no new over-filing.
17 September 2026 · Filing
What changed for you
Drop one source that's about several things — a company, the people in it, its documents, an event — and every agent now files a page for each of them in the same turn. The rule that makes this happen is now one shared instruction every agent carries, not a reminder pasted onto a single agent — so the behaviour is reliable across the board, not just where we hand-tuned it.
What we taught the agent
Every agent now carries one shared instruction about how to keep your team brain — the same for all of them, stated once. In plain terms: keep it complete, clear, and trustworthy so a page can be trusted and acted on without re-opening the original file.
The part that fixed this: complete. A dropped source is usually about several distinct things at once. The agent should give each of them its own page — because a fact filed onto only the main subject is a fact your team loses the moment they look under any of the others. (A name that merely passes by still gets a link, not a page — we didn't trade this for clutter. More on that below.)
What it changed — measured before and after
We took one realistic drop — a partnership proposal naming a company, two people, a document, and a kickoff event (five things worth a page) — and ran it 10 times before the training and 10 times after, changing only this one instruction.
To measure the new shared instruction on its own, the "before" run has our old per-agent note — a hand-written reminder we'd pasted onto one agent to nudge this same behaviour — removed. So the 6-in-10 is what an agent does with neither aid; the 10-in-10 is the new shared instruction carrying it alone. (The one agent that already had the old note was already near 10-in-10 — the point here is that the shared instruction matches it, so the note can be retired and every agent gets the behaviour, not just the one we tuned by hand.)
Page-worthy people filed · out of 10 drops
Fully complete filings · all 5 subjects, out of 10
The two people — clearly worth their own page — were dropped in 4 of 10 drops before, and in none after. The one subject still occasionally left off is the future kickoff event: it's a genuine judgement call (it hasn't happened yet). Our strict "all five" count does still include it — that's why that bar reads 9-in-10 rather than 10 — but the unambiguous subjects (the people) go all the way to 10-in-10.
Why this matters
The whole promise of a team brain is that you can ask it "what do we know about this client?" and get a real answer. That only works if the people, documents, and companies each have a page. This change is that promise holding for a messy, real-world drop — not just the tidy ones.
We checked it didn't start over-filing
Teaching an agent to file more risks it filing junk. So we re-ran our full set of restraint checks with the new training on: 18 out of 18 passed. A price quote still becomes one line on the company's page (not a page of its own), a drop that adds nothing new still files nothing, and an ambiguous "which company?" still asks instead of guessing.
How we trained it — one place, not one agent
The interesting part for the team: this used to be a per-agent workaround — a note pasted onto Snoopy alone. Now it's a single shared instruction every agent gets automatically. The measured result matched the old per-agent note exactly (people filed 10-in-10), but from the shared layer — so the workaround can be retired, and every future agent inherits the behaviour for free.
What's next
- Retire the old per-agent note now that the shared instruction does the job.
- Extend the same "keep it clear and trustworthy" training to how the agent answers from the brain — leading with what's true now and naming what it doesn't know yet.
Training log
A dated log of the big changes to how our agents are trained, each with the measured before/after. Read it to see how the team brain gets better over time.
Connectors
Give agents read-only access to approved GitHub repositories and Linear teams, file and edit Plane work items from a channel, save files to Google Drive and work in Google Sheets, and see merged pull requests as updates in a channel.