Skip to content
AIMUG
← All talks

Mon, Nov 2, 202625:38Coding agentsAgents

Self-Improving AI Harnesses: Latest Research

Share this talkLinkedInX

Julian Ghadially

Julian Ghadially shows agent-directed search matches evolutionary AI harness optimization.

Watch on YouTubeRecorded at AIMUG, Nov 2, 2026

Takeaways

  • Agent-directed search is just as good, if not better, than evolutionary search.
  • Evolution was beneficial on the old models, but it's just not seen anymore with models that are much more capable coding agents.
  • There's literally no difference between the exploration for the Pareto evolution versus the agent-directed search.
  • The starting state for agent-directed search doesn't matter so much, because if it has a bunch of bugs, the agent will usually find those bugs, fix them.
SpeakerJulian GhadiallyCo-founder, CodeEvolverAll talks →

Transcript

Transcript coming.

Same night

6:16AIMUG welcome and announcementsCommunity
Premieres Fri, Nov 620:09Building my first iPhone & Apple Watch app to train for a 5KSal Castoro · Agents
Premieres Mon, Nov 940:41Black October: a security filter switched off by a nameColin McNamara · Inference

Related talks