ALTK-Evolve cuts agent inference tokens while matching or improving ACE
IBM Research authors compare ALTK-Evolve with Agentic Context Engineering (ACE), two systems that turn an AI agent’s previous task trajectories into reusable guidance without changing model weights or using human labels. Both retain individual lessons rather than reducing them to a short summary. ACE maintains a comprehensive playbook, while ALTK-Evolve stores separately retrievable guidelines with support counts and provenance.
The main difference is delivery. ACE adds the full playbook at every agent step. ALTK-Evolve can select a small, task-specific group of guidelines, or provide the full consolidated set when a model can use it effectively. The authors say this calibration reduces inference tokens while preserving access to the same lessons.
In in-house AppWorld tests with the same ReAct agent, DeepSeek-V3.2 reached 89.3% Task Goal Completion and 80.4% Scenario Goal Completion with ALTK-Evolve, versus 80.4% and 73.2% for ACE; average tokens per task were 263K and 634K respectively. With gpt-oss-120b, ALTK-Evolve scored 56.0% and 37.5%, compared with ACE’s 54.8% and 35.7%, using 116K versus 777K tokens. The results were single runs on 168 test tasks.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.