THE NUMBER
More than 130 investors.
That is how many an AI agent fielded questions from while running a $100 million funding round for the company that built it. It drafted investment memos and tracked which slides investors lingered on. The round closed at roughly a $500 million valuation, no founder flying anywhere. Every number there comes from the company itself.
3 THINGS HAPPENING RIGHT NOW
An agent ran a $100 million fundraise
Lyzr, a New Jersey company that builds AI agents, handed its own fundraise to an agent called SivaClaw. It fielded questions from more than 130 investors, drafted memos, tracked which slides held attention. Lyzr says it pulled in $400 million of interest. No reason to doubt them, though TechCrunch wrote "reportedly," because nobody outside the company was in those conversations.
Alberta scanned 466 million lines of code in 20 hours
The provincial government pointed about 50 Claude agents at 3,400 repositories across its 27 ministries. Roughly 20 hours; by hand, an estimated six and a half years. Nobody on that team can personally verify 466 million lines. The record of what the agents did is all anyone has.
Your agent no longer needs your computer to be on
Claude's Cowork tool now runs scheduled tasks with no device online. Not while you are away from your desk. While your machine is off. Whatever happens then, the only account of it waits for you in the morning.
THE DEEP DIVE
When the worker writes the report, the report is not evidence
The Lyzr story is remarkable and I do not doubt it. I want to point at its shape.
An agent had 130 investor conversations. Nobody outside the company sat in. The agent wrote the summary of its own work, the company passed it to a reporter, the reporter wrote "reportedly." Every layer behaved correctly. There is still no evidence in the chain, only accounts.
This is not a Lyzr problem. It is the ordinary condition of handing work to something that also narrates what it did. Any owner who has heard "I called them all, nobody picked up" knows the feeling, except with a person you have a phone bill. With an agent, the summary comes from the same thing that did the work, so it inherits every blind spot the work had. A skipped step does not appear as "I skipped a step." It does not appear at all.
Absence is invisible in a self-report. That is the whole problem, and it is about to be yours.
ONE THING TO TRY THIS WEEK
Run the same instruction twice and read the difference.
I promised you a way to see what your agent did rather than what it told you. Asking it will not work: it can replay steps you could scroll up and read, but its judgment calls were never recorded, so it reconstructs a plausible answer and you believe it. Here is the trail you can actually get. Run the instruction again in a fresh window, and the decisions it made for you come back different.
Find a real instruction you gave an AI recently. Copy your exact words.
On paper, write the one thing you are surest it got right. "It only contacted people from this year." "It used my pricing." Whatever you would bet on.
Open a brand new chat in whatever tool you use and paste that same instruction. Nothing added, no mention of the first attempt.
Put the two side by side. Ignore wording and length: it phrases things differently every time and that means nothing. Look for differences in what it did. Did one include people the other left out, use a different date range, quote a different price? Anything that changed is a choice you left open.
Check your paper. If the thing you were surest about came back different, that is the one to fix, and you would not have found it in either answer alone. If it held steady, you have confirmed something you can trust it on.
Rewrite the instruction with those differences nailed down. That version is safe to hand over.
Stuck? Reply to this email. I'll help.
WHAT'S COMING
Next issue: what happens when you put a second agent in the room whose only job is to find fault with the first one's work. It catches things you never would.
Manu