Gehan Panapitiya
Agent memory & self-improvement engineering lead · Independent researcher · Christchurch, NZ
The companies that thrive in the AI-native era won't be the ones that automate work by adding more agents to production. They'll be the ones that make self-improvement stable — so improvement can be pointed at any dimension of the business, and the value arrives now rather than after another quarter of people writing agent code.
Stable self-improvement is also what makes AI governance enforceable. You cannot govern a system you cannot keep stable, and you cannot keep humans safely in the loop with agents whose memory drifts between runs.
I work on that problem from two sides. In practice, I help engineering teams get long-horizon agents stable and under control in production — memory, evaluation, guardrails — whatever cloud or model stack they are on. In research, I study recursive self-improvement and long-horizon memory: how an agent that rewrites its own capabilities keeps a stable, trustworthy memory across long runs and across a swarm, and what has to decay for that to hold.
Until October 2026 I led Platform Engineering at Phocas Software. Before that: Xero, Diligent, Tait Communications, HSBC.
Problems
The hard problems in stable self-improvement and swarm memory, as I see them from production. One problem per piece; no solutions until the problem is clear.
First pieces coming shortly.
Research
Preprint in preparation.
Conversations
Recorded conversations with practitioners who have hit these problems — one named problem each, twenty minutes, no product.
First recording to follow.
Work with me
I take on contract and advisory work where a team is stuck getting agents stable in production. If that is you, say so on LinkedIn or X.