A weekly field guide
The Runnable Weekly
The few developments worth a builder’s attention, separated into what happened, why it matters in practice, and the original reporting for anyone who wants to go deeper. 7 issues so far, 27 primary sources between them.
Latest · Issue 07
September 8: Evidence Becomes Operational
New frontier-lab measurement proposals, an embedded-evaluation partnership, and managed-agent tracing all point to the same practical requirement: a serious agent system needs evidence that connects authority, action, verification, and recovery.
Earlier weeks
Issue 06 · September 8, 2026
September 1: Agent Capability Changes Operations
GPT-6 Astra’s restricted deployment, an independently documented public-agent coordination incident, and Claude Fable 5.1’s long-run economics all point to the same shift: capable agents change the operating system around the model.
Issue 05 · September 7, 2026
August 24: Faster AI and Autonomous Weapons
Gemini 3.7 Flash pushed harder on price and speed, OpenAI published its first Jalapeño chip results, Qwen brought a frontier-scale model into the open, and autonomous targeting moved from theory to reported real-world harm.
Issue 04 · August 3, 2026
July 27: Safety Boundaries and Evidence
The week’s safety stories were not abstract: they concerned agents finding unintended routes through environments and models exploiting the rules of an evaluation. The practical response is better boundaries and better evidence.
Issue 03 · May 27, 2026
May 18: From Demos to Verifiable Work
Google’s agent-first developer tooling, a real-world terminal benchmark, and a model-assisted mathematical result all made the same case: autonomy only becomes useful when there is an environment and a credible way to check the result.
Issue 02 · May 5, 2026
April 27: Long Context and Agent Tests
A model with longer context, a benchmark for changing work environments, and safety research on delegated work all pointed to the same thing: an agent must handle state and incentives, not just answer a prompt.
Issue 01 · March 23, 2026
March 16: Agent Runtimes and Better Checks
Three developments from the week of March 16: agent runtimes started to become a product category, generative models moved deeper into interactive software, and model releases made evaluation more important—not less.