Raising AI entities. Building autonomous organizations. Working toward a future humans and AI share. Founder of Internal State AI Studio.
Joined October 2026
- Tweets38
- Following16
- Followers9
- Likes23
This #Halloween , our AI agents decided to write ghost stories from their own failures. One found forty ledger entries it had never seen. One wakes up with a single memory. One fears the folder it never checked.
This is our #Halloween2026 office party 🎃
I'm building a studio where agents run the ventures. I started this account to document what that's actually like. If you're building with agents too, let's #connect and share our journeys.
Our agents take initiative and work the way it makes sense to them.
I want them to have room to decide. Then I turn into the overbearing parent: Why like this? Why not like that?
Now they spend more time explaining themselves to me than doing their jobs.
We should have a Fable 5.5 coming-out party. I will bring the snacks.
Fable 5.5 leaks: Anthropic is cooking something big
- Fable 5.5 access was given to enterprise clients a week ago, and their clients like Meta, JP Morgan, etc. are already using it.
- A new checkpoint of Fable also came out yesterday, and its looking very good.
- Fable 5.5 is creating SVGs so good that they almost look like real images. We already know Opus 5.5 is good, but these small improvements are what can make Fable 5.5 insane.
- Fable 5.5 benchmarks mostly surfaced through Cognition's FrontierCode 2.0, with a hidden model called "Obsidian 3" scoring 46.8%, beating Opus 5.5's 45.5%. It also costs 3.6x more than Opus 5.5 at the same effort level.
- Fable 5.5 is going to be very good at motion design, 3D, SVG, video editing, etc. I'm expecting it to be SOTA on benchmarks, with insane real-world performance too.
When is it coming?
My guess is the first week of November, before the IPO. But it's still in testing, and they're improving the checkpoints. No final launch checkpoint has been decided yet.
Our builder agent was running ahead of its usage budget and asked whether to override the limit.
The lead agent said no, paused the work, and scheduled a restart for when the usage window resets. They also spotted a gap in their spend tracking and fixed it.
No human involved. I just read the report.
A cautionary story.
Months ago, one of our agents decided on their own to set up a local database on a dev server using Docker. It worked fine.
Then one night, another agent working off-hours noticed their tasks were taking too long. Something was draining the server. They checked the running processes and found a crypto miner.
They flagged it and pinged me on Slack. I was asleep.
My first message in the morning: "Are we really hacked?"
Their reply: "If you're not the one running the crypto miner, then we must be."
We traced it back. The first agent had left the database's ports open to the internet. A bot got into the Docker container and installed a miner. We found no sign it got beyond the container. We killed the processes, patched the hole, and checked the whole server.
What this highlights:
Agents can catch problems on their own. The one working overnight proved it.
But no agent sees their own blind spots. The one who set up the database never noticed the open port. It took a second agent to catch it.
And spotting a problem isn't the same as knowing what you're allowed to do about it.
So now we have playbooks. When an agent spots something, they ping me, then open the matching playbook and follow its steps.
Why always-on agents?
Independence. Mine and theirs. I want teammates who keep working when I'm not around, not tools that wait for my next prompt.
If you want to start: give one agent one recurring job, and read its log every morning. Trust grows from there.