This overview covers 2026-07-13 through 2026-07-19 in Asia/Taipei time. W29 in one sentence: base models still matter, but the external harness increasingly determines whether an agent is affordable, reliable, safe, and able to improve.

The three deep dives

• Industry: the operational runtime wins (Grok 4.5, AgentCore, Qwen Cloud)
• Research: harnesses learn—and need fair tests (MemoHarness, E3, GPT-Red)
• Open source: harness engineering becomes product (Agent Framework Go, CORAL, Browser Use)

What the three tracks say together

Industry packaged cost, isolation, voice, MCP, and deployment as runtimes. Research explored learning from experience and scaling effort to task complexity while warning that benchmark search can masquerade as generalization. Open source turned middleware, checkpoints, observability, and self-evolution into frameworks.

Their intersection is the control layer as core asset. The next leaders must answer: what did the agent see, why did it act, what did it cost, how did it recover, and which experience changed its next behavior?

Builder and operator guidance

Freeze the harness and acceptance set when comparing models. Record success, tokens, wall time, tool errors, human handoffs, and security events together. Put least privilege and confirmation around sensitive tools. Constrain self-improvement with held-out tasks, versioned policy, and rollback.

Watchlist

• Full cost per successful task, not token price alone.
• Harness learning that transfers without accumulating bad memory.
• Third-party reproduction of automated red teams against new tool surfaces.
• Stable releases, migrations, and secure defaults in open-source frameworks.
• Public incident data for browser and voice agents.