Latest issue ·
Issue #2
Covers October 6 – November 5, 2026
Running agents for real: oversight that holds up, harnesses that make cheap models reliable, identity and failover, and how work changes inside an AI-first company.
This issue is about what it takes to run agents for real. Anthropic shows both sides: experts quietly deferring to their agents, models working around restrictions on the open web, and a company where agent hours now outnumber employees six to one. On the engineering side, a harness experiment makes a cheap model as reliable as a frontier one, and Google argues for measuring cost per successful task. If you read two, start with the harness experiment and Anthropic’s report on how employment is changing.
In this issue 13 posts
- Agents aren’t humans (but they kinda are) Cloudflare: AI
- Cost per task: Using cheap classifiers to route across Gemini models and thinking levels Google Developers: AI
- Why are good developers so wary of agentic development? Martin Fowler
- Le Chonk: Built for Coders, Trusted by Defenders. Mistral AI
- How AI agents shape experts' judgment calls Anthropic Research
- The five components of building a "company harness" LangChain
- LLMs are making formal software verification feasible at a much larger scale Anthropic Research
- How To Use Decision Models for Evals Hamel Husain
- How employment is changing at Anthropic Anthropic Research
- Designing failover across OpenAI model providers OpenAI Developers
- Redesigning source control at Anthropic Anthropic Engineering
- Are Harnesses Worth the Cost? Martin Fowler
- Investigating unintended model actions in our evaluations and internal use Anthropic Research