A closing batch about where the work, and the money, is heading, from the codebase out to the economy, including whether small software businesses survive AI.
Addy Osmani’s The New Software Lifecycle pulls the ideas that matter out of a Google whitepaper he co-wrote. The framing I keep reusing: an agent is a model plus a harness, roughly 10% model and 90% harness (the rules, tools, sandboxes, orchestration, guardrails, observability), so when an agent does something dumb, debug the harness first, because most agent failures are configuration failures you can fix today. Context engineering (static vs dynamic) is a financial lever, not just a technical one; verification (tests plus output and trajectory evals) is the line between vibe coding and engineering; and the lifecycle compresses unevenly: implementation collapses to hours while specification and verification stay slow, so spec quality becomes the bottleneck. “Set the bar at the eval, not the demo” is the sentence to hand a skeptical lead.
Armin Ronacher’s The Coming Loop is the uneasy counterweight, and, full disclosure, his third appearance across this series, so I’m flagging it (more on that in the dedup note; you may want to space him out). He’s watching harness-level loops (a queue, a machine that attempts work, a harness that decides whether it’s actually done and keeps it alive past “I’m done”) take over agentic engineering, and he’s honest that they already work astonishingly well for porting, performance exploration, security scanning, anything that transforms existing code or produces short-shelf-life artifacts. His worry is lasting code you care about: loops amplify the model’s instinct to add local defenses instead of making bad states impossible, and the deeper fear is dependency: codebases produced, reviewed, and maintained by loops that assume machine participation, that you can no longer understand without the machine. And you can’t fully opt out, because attackers and competitors will loop against you. It pairs naturally with the Osmani piece: same future, opposite temperament.
Brandur Leach’s The Minimum Viable Unit of Saleable Software is the SaaS-economics read, and a genuinely useful mental model given your own product ambitions. Starting from the viral “we replaced $400/mo Jira with a Claude-built tracker” anecdote, he does the actual math: cheap isn’t zero, LLM-built software still costs feedback loops and maintenance labor, and the Jira clone takes ~37 months to break even while a $25k/mo Salesforce is closer to a real build decision. He proposes a “zone of viability” (enough novelty that a rebuild is non-trivial, priced low enough not to provoke one) and the “minimum viable unit of saleable software” below which buying beats building. He’s betting his own job queue business on landing above that line. Required reading before you price anything.
Finally, Hanna Horvath’s I Make Good Money. Why Do I Still Feel Like This? zooms all the way out to the economy the rest of us are building software inside. It’s a sharp essay on the K-shaped economy and the “missing middle”: the American middle class as a deliberate postwar policy project that got unbundled and repriced piece by piece, splitting into “material precarity” (the basics slipping) and “positional precarity” (high earners whose expectations no longer resolve). The real line, she argues, isn’t income but capital (whether your money generates its own returns), and both groups misdirect their frustration sideways instead of up. US-centric, but the psychology of the treadmill travels. A good, humane palate cleanser after a lot of AI discourse.
End of batch.