◆ Selected work
Self-improving agent platform
The agentic system I actually run every day. It mines its own sessions for durable lessons and writes them back into memory, and earns new capabilities through single-purpose allowlisted scripts rather than wider permissions.
In daily use
A production agentic setup I use daily, not a demonstration. Two parts of it are worth describing because both were arrived at the hard way.
The first is a self-learning loop: it reads back its own sessions, extracts what turned out to be durable, and writes that into long-term memory. The discipline is in what it refuses to save — anything unverified, anything private, anything that only mattered once.
The second is how it gains capability. When the agent needs to do something it is not permitted to do, the answer is a single-purpose script that is allowlisted by name, never a broader permission. Narrow and auditable beats convenient.
The repository is private, so there is no link here. Happy to walk through the design.
No comments yet — be the first to add to the discussion. Comments appear after they’re reviewed.