The New Delegation Stack

Sun, 09 Aug 2026 00:00:00 +0000

Somewhere in Anthropic’s production stack there is a prompt that teaches one of the most capable models on Earth how to be a middle manager. It spells out staffing policy by hand: simple fact-finding gets one agent and three to ten tool calls; a direct comparison gets two to four subagents with ten to fifteen calls each; complex research gets ten or more subagents with clearly divided responsibilities (How we built our multi-agent research system). They wrote those rules because, without them, the lead agent did what every new manager does — early versions spawned fifty subagents for simple queries (the review’s account of the system).

Evals Are the New Engineering Artifacts

Sun, 31 May 2026 00:00:00 +0000

Traditional software relies on explicit specifications. AI systems increasingly require a different contract: the ability to measure whether behavior is good.

Evaluation becomes a core engineering artifact.

It defines what success means, creates feedback loops, and allows intelligent systems to improve.

Thoughts on Sina's Page

The New Delegation Stack

Evals Are the New Engineering Artifacts