---
title: "Huawei openJiuwen adds TTSE framework for AI agents"
slug: "huawei-openjiuwen-adds-ttse-framework-for-ai-agents"
date: "2026-11-11"
category: "AI & Automation"
author: "Dr. Kai Wei"
correspondent: "Dr. Kai Wei"
correspondentRole: "Lead Frontier AI Correspondent"
readingTimeMinutes: 2
tags: ["AI","News"]
originalSourceUrl: "https://pandaily.com/huawei-openjiuwen-ttse-two-track-self-evolution-agent-experience-fact-tip"
canonicalUrl: "https://aurav.ae/huawei-openjiuwen-adds-ttse-framework-for-ai-agents"
publisher: "Aura-V Intelligence (aurav.ae)"
---

# Huawei openJiuwen adds TTSE framework for AI agents

> Huawei’s openJiuwen platform introduces TT SE, a two-track system that separates factual environment data from conditional tips and reports benchmark gains.

## Key Takeaways (TL;DR)

- TTSE splits agent knowledge into FACT and TIP, improving learning efficiency.
- Benchmarks show 97.8% success on ALFWorld and a 30.09% reduction in output-token use in office demos.
- Available in WorkSwarm 0.2.7+, users can enable TTSE via a setting.

Huawei’s open-source agent platform openJiuwen has introduced TTSE (Two-Track Self-Evolution), a framework that categorises learned information into environment facts (FACT) and conditional tips (TIP). The rollout is described in an arXiv paper submitted on September 21 and was reported by QbitAI on November 10 [QbitAI Report](https://www.qbitai.com/2026/11/508998.html).

TTSE starts with an empty rule base, induces initial rules from a first batch of tasks, then iteratively refines them by retiring misleading rules, merging contradictions into TIPs, and retaining retired rules for traceability.

OpenJiuwen’s own tests on five benchmarks report the following results: ALFWorld reached 97.8% success with DeepSeek-V3.2 (up from 83.6% for GraSP) and 98.51% with GLM-5; ScienceWorld rose from 53.08% to 73.46%; GDPevo scored 0.2616 versus 0.0947 for zero-shot; SOPBench achieved 94.58% success versus 79.15% baseline; PinchBench improved average score from 0.7439 to 0.7822.

In two office demonstrations, an Excel data-cleaning task dropped from 57 tool calls and 6 min 22 s to 20 calls and 1 min 48 s after TTSE experience injection; a document-and-table task fell from 27 calls to 18. On the WorkBuddy Office Bench, openJiuwen reports a 3.81% performance gain and a 30.09% reduction in output-token use.

TTSE can be enabled in WorkSwarm version 0.2.7 or later; the article does not confirm a standalone library release.

Our read: TTSE demonstrates a concrete step toward modular agent cognition, with measurable gains across diverse tasks, suggesting Huawei’s platform may become a more competitive option for enterprises seeking efficient, self-optimising AI assistants.

---

*Original Think-Tank Source: [Huawei's openJiuwen Unveils TTSE, an Agent Framework That Splits Environment Facts From Tips](https://pandaily.com/huawei-openjiuwen-ttse-two-track-self-evolution-agent-experience-fact-tip)*