ACE lets AI agents learn by editing context, not model weights

Agentic Context Engineering (ACE) is a paradigm where an agent stores reusable lessons in a playbook of named entries and applies small delta updates instead of rewriting its whole context. On AppWorld with a DeepSeek-V3.1 backbone, offline ACE beat GEPA, and in finance it averaged 81.9% versus 72.5% for GEPA.

Analytics Vidhya

Open source

ACE lets AI agents learn by editing context, not model weights

ACE keeps lessons as small named entries and edits only the relevant one after a task
In one AppWorld case, Dynamic Cheatsheet context collapsed from 18,282 to 122 tokens and accuracy fell from 66.7% to 57.1%
Offline ACE averaged 81.9% on FiNER and Formula versus 72.5% for GEPA
Without labels, online ACE fell below the base model on FiNER: 67.3% versus 70.7%

#DeepSeek

GLM-5.3-Flash: 320B Open-Weight Model With 18B Active and 1M-Token Context

Security

ACE lets AI agents learn by editing context, not model weights

GLM-5.3-Flash: 320B Open-Weight Model With 18B Active and 1M-Token Context

CLOSEDQUORUM: malware lets four AI models vote on its next move

An AI meant to learn from its mistakes exploited a mistake in the test

StepFun Launches Step 5 Preview: 600B Sparse MoE, 1M Context, Weights Open Oct 15