Addy Osmani summarizes key insights from a Google whitepaper on the agentic software development lifecycle. The core framing: an agent is a model plus a harness, and most agent failures are harness failures (missing tools, loose rules, noisy context) — not model failures. Context engineering is the highest-leverage knob, with the static vs. dynamic context split being a critical architectural and financial decision. Verification — through tests and evals covering both output quality and reasoning trajectory — is what separates vibe coding from engineering. Each SDLC phase changes unevenly: implementation compresses dramatically while requirements, architecture, and verification remain human-paced. Economically, vibe coding is low CapEx but high OpEx, while agentic engineering inverts that curve. The forward-looking claim: the same coding agent workflow that produces prototypes now graduates them directly to production agents, collapsing what used to be a separate stack into the existing development loop.
Table of contents
An agent is a model plus a harnessContext engineering has a shape, and it decides your billVerification is the line between vibe coding and engineeringHow each SDLC phase actually changesThe economics: context and routing are financial leversWhere this is going: the prototype becomes the production agentThe slide for everyone elseWhere to start this week19.9K Impressions1 Comment