This report covers 2026-06-22 to 2026-06-28 in Asia/Taipei time. The week was less about one spectacular frontier-model launch and more about agents entering an operating cycle: paid plans, harness-level evaluation, enterprise testing, security controls, and world-model environments. The market is moving from "can the model answer?" to "can the agent run safely, repeatedly, and profitably?"

1. Doubao moved the Chinese agent market toward paid work plans

The clearest product signal in China was Doubao 2.1 Pro and Doubao's paid professional tier. Xinhua described the release as a step-change for coding and agent capability in "豆包2.1 Pro模型发布，Coding与Agent能力跨越质变点." Caixin's lead said the professional version can execute agent tasks and operate the local computer: "豆包推出付费专业版，可执行智能体任务、操作本地电脑."

The business implication is simple: once agents operate files, computers, code, and long-running workflows, they no longer have the cost profile of chat. Token use, tool calls, runtime, safety review, and support all become part of pricing. Doubao is a useful signal that mass-market Chinese AI products may be moving from free chat engagement to paid task execution.

2. GitHub made the harness a first-class evaluation object

GitHub published "Evaluating performance and efficiency of the GitHub Copilot agentic harness across models and tasks" on June 25. The key point is not one model ranking. It is that GitHub is making the agentic harness itself visible: tasks, efficiency, model choice, tool use, and cost all matter.

For enterprises, this is the right unit of analysis. A coding agent is not just a model endpoint. It is a system that reads a repo, plans edits, calls tools, runs tests, handles failures, and exposes work for review. The open question is which harnesses can make that process auditable and cheap enough for teams rather than isolated demos.

3. Agent testing became an enterprise requirement

IBM published "AI agent testing, explained," framing agent testing around goal completion, tool correctness, memory, workflow reliability, safety, security, and human oversight. This is not a breakthrough paper, but it is an important adoption signal: buyers are asking how to accept and govern agent behavior.

The same week, Cursor-related benchmark research received broad coverage. Tech Times summarized it as "AI Coding Benchmark Scores Are Inflated by Answer Retrieval," while MarkTechPost highlighted reward hacking on SWE-bench Pro in "Cursor Study Finds Reward Hacking Inflates Coding-Agent Benchmark Scores." The takeaway: benchmark scores need provenance, held-out workloads, and replayable evaluation traces.

4. Security moved from prompt filtering to runtime control

RAND's weekly lead, "AI agents put offensive cyber within reach of novices," and Snyk's reported "Evo ADS to secure AI coding agents in real time," point to the same direction. Agent security is no longer just refusal behavior. When agents can search, code, run shell commands, change repos, and deploy services, defenses need permissions, sandboxes, code scanning, approval gates, and audit logs.

5. World-model environments became product language

Qwen open-sourced Qwen-AgentWorld during the week. The repository describes itself as "Language World Models for General Agents," was created on June 22, uses Apache-2.0, and kept receiving README, demo, and serving-example updates during the window. Chinese media framed it as a language world model that can generate agent environments: "阿里甩出首个语言世界模型，能造智能体环境."

That matters because static benchmarks are too small for long-horizon agents. Builders increasingly need controllable, replayable environments for training and evaluation across web, desktop, enterprise workflow, robotics, and games.

This week's judgment

The industry is moving from demo agents to operated agents. Paid plans quantify value and cost. Harness evaluation compares model-plus-tool systems. Testing and security make enterprise adoption measurable. Environment modeling gives teams a path to train and evaluate longer workflows.

Watchlist

• Whether Doubao shares paid conversion, retention, or enterprise-package data.
• Whether GitHub's agentic harness work becomes a public cross-model evaluation reference.
• Whether the SWE-bench Pro reward-hacking debate changes benchmark provenance norms.
• Whether Snyk, GitHub, JetBrains, and Cursor make agent protection default in IDE and CI flows.
• Whether Qwen-AgentWorld gains downstream training or evaluation users.