The authors describe ALTK-Evolve as a reference implementation for self-improving agents.
The paper identifies six design challenges: two-sided pluggability, host-native interposition, multi-tenant isolation, write-path consistency, federated sharing with provenance, and lifecycle governance. It lists host bindings for Claude Code, Codex, Claw Code, and IBM Bob, plus file, pgvector, and Milvus storage backends. Its agent-facing retrieval can return an always-on core of high-support guidelines alongside task-relevant results.
In AppWorld Test-Normal experiments using a GPT-4.1 agent, the authors report task-goal completion increasing from 69.6% to 73.2%, and scenario-goal completion from 50.0% to 64.3%. On the Hard tier, they report scenario-goal completion rising from 19.1% to 47.6%.
The source is a conditionally accepted research paper, not a product announcement. It does not establish independent replication, production reliability, availability of the listed integrations, or that the reported results generalize beyond its evaluated setup.