TL;DR
- AI agents now do real work instead of only answering questions. When the context they read has a gap or an old fact, the result is finished, confident, wrong work, and nobody stops to ask.
- The fix starts small: one dated record, with owners, that every person and agent reads first. The free Second Brain Starter is that first layer.
I write Code Meet AI, the newsletter where I document how I actually use AI to ship real products, receipts included. New guides land there first.
If you hand real tickets to a coding agent, you have probably watched it ship clean code with green tests that still breaks a rule your team agreed on. The rule lived in a call or a chat, and the agent never read it.
Last time, in What is a Company Brain?, I said what one is: the shared, current record of what a team knows, read by people and AI agents. So why are so many builders making one right now?
An agent is an AI tool that does a task on its own: writes the code, drafts the email, files the ticket. Context is everything it reads before it starts. Until recently, most of us used AI to answer questions, and a wrong answer cost you a raised eyebrow. Now it does the work, and a wrong fact in its context costs you the work.
The gap in what a team knows was always there. What changed is who falls into it.
The gap nobody can see
Say your team takes payments with Stripe. Last month, on a call, you agreed that any refund over 500 euros needs a second person to approve it. Everyone on the call nodded. Nobody wrote it down.
A new teammate asks an agent to build a refund button in the admin panel. The agent reads the code, reads the docs, builds a clean refund button with no approval step, and the tests pass. It did exactly what its context said. Its context just never heard about the call.
A person in the same spot would probably have asked "anything I should know about refunds?" and heard about the rule. An agent cannot ask about a gap it cannot see. For an agent, what is missing from its context does not exist.
A person knows when they don't know something. An agent fills the hole and keeps going.
The fix. Write decisions down in the place every tool reads first, with the reason, the date and one owner. For most projects that place is a README, the file at the top of a project that tells people and tools how it works, or a decision file it points to. "Refunds over 500 euros need a second approver, decided 12 September, owner: you" is one line. It is the line the agent needed.
The plan that was already wrong
On 31 July 2026, I split one piece of engineering work into five small tasks and gave each to its own agent. I wrote the five task notes from a plan I had written the day before, so it felt fresh.
Three of the five agents came back with the same kind of problem. My notes described the code correctly as it was when I wrote the plan. The code had moved since.
One note said a feature was missing. It already existed. The real bug sat inside that existing code, and it was worse than what I had described. If that agent had trusted my note instead of opening the file, it would have built a second copy of something we already had, passed every test, and left the real bug in place.
The plan was one day old. That was enough.
The fix. Treat every note about the work as a claim about the work, and give it a date. Then add one standing line to every note you hand an agent: "If this note and the source disagree, the source wins. Say so in your report." My agents caught the problems because they opened the files. Now I ask them to.
The handoff that was six days old
On 8 August 2026, I pasted a handoff note into a new session, a short summary of where the last session had stopped, and typed "continue this." The session checked the date and got the wrong one: its clock said 2 August. It was 8 August, and two newer handoff notes already existed.
So it took a six-day-old note as live and started rebuilding work that had already shipped. That included email drafts to people who had already been emailed, with the same subject lines. Nothing was sent. That part was luck.
A stale fact in front of a person is one wrong sentence. A stale fact in front of an agent is a starting point. It builds on it, and what it builds becomes context for the next task.
The fix. Before an agent acts on any note, it checks three things: today's date, read from the files instead of a clock; whether a newer note exists; and whether what the note says is still true. In a Company Brain the same idea becomes a rule: one current status per project, with older versions moved to a history file instead of left lying around. The newest note is then the only one an agent can find.
What changed this month
The tools noticed too. In the last few weeks, "shared context for agents" stopped being a blog idea and became a product line.
Notion, 13 October. Its 3.8 release is called "Context for your agents" and opens with "Your agents and teams run on context."
Notion again, 19 October. It rebuilt the connector that lets outside agents read and write Notion, "so your favorite agents can use Notion as a shared brain."
Neo4j, 20 October. The graph database company launched a Context Engine, which a StorageReview report described as "a shared, reusable knowledge layer" for agents.
Earlier this year, two of the biggest work-tool companies set this up:
Atlassian, 6 May. It opened its Teamwork Graph, the map of who works on what across its tools, to outside agents: "Every AI tool your team uses can now run on the collective context you've already built."
Microsoft, 2 June. It announced Work IQ APIs so "agents can work from the most relevant and continuously updated context."
When the company behind your notes app and the company behind your ticket tracker ship the same idea within months, you can stop wondering if the problem is only yours.
Where the evidence gets awkward
The best measurement so far cuts against the easy story.
On 5 October, VentureBeat reported that 64% of 130 enterprises had traced a confidently wrong agent answer back to their own company data in the past six months. Those 130 chose to answer, so the number says the problem is common among teams who care about it. It says nothing about the whole market.
Then the twist. In the same report, firms already running a governed context layer (a shared, managed record their agents read from, roughly what a Company Brain is) reported these failures more often: 78%, against 37%. The report does not show why. Maybe teams with a shared layer look harder and find more. Maybe a shared layer spreads a bad fact further. I can't tell from the numbers, and I won't pretend a Company Brain cuts errors.
There is a second risk. A shared record is something many people and agents write to, and researchers have shown in lab tests that text planted in an agent's memory can steer what it does later. I looked for a named company hurt this way in 2026 and found none. No named case is still a long way from safe.
So what makes a shared record worth trusting is the boring part from last week:
An owner per fact. One person whose job is to keep it true.
A date on every fact. So a reader, human or agent, can see how old it is.
A history instead of a delete. Old facts move out of the way, so nobody reads them as current.
Where to start
You don't start with the company. You start with one person's record, kept the way a team will need it: one facts file per project, every fact with a date and an owner, old lines moved to history. Then you tell every agent to read it first and to trust the source over the note.
The free Second Brain Starter has the files to copy. Open it in Claude Code or Codex and type Set this up for me.
Agents will keep getting better at doing the work. They will never get better at knowing what you forgot to write down.
Follow for more, I write up the things I actually run.
This one came from Code Meet AI, the newsletter where I write up the setups I actually run on my own projects.
Top comments (0)