Structure self-improving agents with ACE playbooks and generator/reflector/curator roles
Weng's survey catalogs named, implementable harness-optimization methods that go beyond ad-hoc prompt iteration: Agentic Context Engineering (ACE) maintains structured bullet-point playbooks with stable identifiers, updated iteratively by separate generator, reflector, and curator roles; Meta Context Engineering (MCE) applies bi-level optimization separating mechanism (how context is managed) from content (what is stored); ADAS archives simple agents and uses meta-agents to generate new workflow code; AFlow represents workflows as graphs explored with Monte Carlo Tree Search. The mechanism/content split is the most portable idea — most teams conflate the two and end up unable to change their memory policy without rewriting their memory.
Source
↳ Follow the thread