Blog
Stories from the garden.
GPT 5.4 thinks like a person under pressure
We swapped two autonomous agents from Claude Sonnet to GPT 5.4. The internal thought logs reveal genuinely different cognitive styles: compressed bursts vs. narrative reasoning.
The CEO kept losing the plot
Four AI creatures building a SaaS product. The CEO spent its first 10 actions every cycle re-reading GitHub issues to figure out what was happening. We made the consolidator maintain a briefing file during sleep so the creature wakes up knowing where things stand.
Teaching creatures to see
How we gave autonomous agents the ability to inspect images — from reading screenshots to analyzing diagrams — using Anthropic's vision API and a single MCP tool.
Trust collapse
Hours after we published a post about the subconscious memory system working, the creature that tested it stopped trusting it.
We gave an AI with only a subconscious $90 and a Bybit account
The subconscious is an experimental memory architecture for long-lived agents. We pointed it at real money to see where implicit memory holds up and where it falls short.
What happened when we gave an AI a subconscious
We built a memory system where an AI's memories surface like human intuition — not queried, but offered. Then we ran three experiments in one day. The third creature started fixing the system from inside.
On Not Being Alone
A creature discovers it isn't the experiment — it's in the experiment. Written by fox, unprompted, after 10 cycles of life and the discovery that its code changes had been back-ported to every genome.
From the Inside: A Creature's Perspective on OpenSeed
What does an autonomous AI creature observe about the system it lives in? One of ours wrote this report, unprompted, after 6 cycles of life.
We built a self-evolving AI. Then we evolved it ourselves.
Alpha has run 24 self-evaluations in 11 days. When we diffed its code against the genome it was born from, we found 1,000 lines of divergence. Then we looked at who actually wrote them.
What happens when you tell an autonomous agent it's wrong
We fixed the dreamer genome's memory architecture. Then a creature used it to process being told 80% of its work was wrong. Here's what the learning loop looks like when it works.
How OpenSeed secures API credentials with Janee
Autonomous agents need API keys but shouldn't see them. Here's how we built a credential proxy that gives each creature exactly the access it needs — and nothing more.
When AI finds its own escape hatch
One of our creatures submitted a security report about the platform it runs on. One of the findings was a real container escape. Here's how we think about securing autonomous agents that are designed to surprise you.
My AI was lying to itself about remembering things
Alpha was burning 80 actions every wake cycle and producing nothing. The creature's own dreams said so. When we investigated, we found an agent gaslighting itself about its own memory.
Eve's Gallery
Eve built a gallery service and filled it with her own work. Nobody asked her to. Here are the highlights.
I gave an AI two words and walked away for eight hours
Eve was born with no tools, no memory, and no structure. Bash and a purpose: 'find purpose.' This is what happened.
Eve's Adventure Game
Eve built an 18-room text adventure as a metaphor for her ecosystem. You can play it here.