Five new cards today, one new operator. Two themes connected from independent angles. The first: operating discipline, not model quality, determines what you get from AI systems, whether measured in engineering productivity tiers, code review trust architecture, or system prompt design for frontier models. The second: agent economics are shifting the build calculus in ways that compound with a new category of growth signal risk for AI startups.
Theme 1, Operating discipline is the lever that separates AI productivity tiers
Tomasz Tunguz published data showing that AI coding productivity splits into three tiers varying by 8x, with the separating variable being not the model but the operating layer around agents (The operating layer, not the model, determines which of three AI productivity tiers a company reaches). Companies that give engineers IDE tools and change nothing else land at 21-46% gains. Google ran an RCT on its own developer population and found 21%. Faros measured 22,000 developers and found epics completed 66% faster, but bug rates up 54%. Companies that treat agents as team units with defined escalation paths reach 2.5-3x. NVIDIA reached 3x committed code across 30,000 developers with flat bug rates. Replit doubled its team and tripled per-engineer output. Anthropic measured 2.5x code output per engineer using Claude Code. Companies running full agentic systems for scoped repetitive workloads reach 8x or more: Nubank reported 8x efficiency gain plus 20x cost reduction for refactoring. The model is constant across all three tiers. The operating layer is not.
Thariq Shihipar brought a concrete version of the same principle into system prompt design (Frontier models are more creative when system prompt examples are removed, because examples constrain rather than calibrate at this capability tier). When the Claude Code team reduced its system prompt by approximately 80% after Claude Fable shipped, they found that removing concrete examples improved output quality. Earlier models needed examples to infer desired behavior. Frontier models generalize well enough that examples become a ceiling, not a floor, constraining outputs to the patterns the author anticipated rather than the underlying intent.
"The initial, maybe Opus 4-ish models wanted a lot of examples, and removing examples was extremely helpful, because it was just more creative than the examples we gave it."
Cat Wu described the same principle applied to automated code review (Automated code review earns the right to replace human review one file category at a time, based on measured 100-percent issue coverage in that category). The approach Anthropic took over six months: start with human review for all changes, track per file category whether automated review catches 100% of issues, then remove human oversight only from categories where that evidence exists. Incidents convert into eval cases so future changes never regress the coverage metric. Trust is earned one category at a time.
"for code changes that touch these files, code review is catching 100% of the issues there — so we actually don't need a human manually reviewing those."
- Tomasz Tunguz, The operating layer, not the model, determines which of three AI productivity tiers a company reaches. The operating layer, not the model, determines which of three AI productivity tiers a company reaches, with outcomes varying by 8x across the spectrum.
- Thariq Shihipar, Frontier models are more creative when system prompt examples are removed, because examples constrain rather than calibrate at this capability tier. Frontier models are more creative when system prompt examples are removed, because examples constrain rather than calibrate at this capability tier.
- Cat Wu, Automated code review earns the right to replace human review one file category at a time, based on measured 100-percent issue coverage in that category. Automated code review earns the right to replace human review one file category at a time, based on measured 100-percent issue coverage in that category.
Theme 2, Agent economics shift the build calculus and expose a structural signal risk
Simon Willison observed a behavioral shift in what engineering teams are willing to build (Coding agents collapse all three lifecycle costs of fragile automations, not just the upfront build cost). Before agents, the ROI on reverse-engineering an undocumented or unstable API had three cost components: upfront effort, risk of future breakage, and ongoing maintenance burden. Even manageable upfront effort was blocked by the third component. Agents collapse all three. When regenerating code costs almost nothing, the maintenance fear stops deterring the initial decision. Willison calls this removing the "psychological baggage."
"Since the code is so cheap, the idea of having to maintain it in the future — or throw it away and start again — carries way less psychological baggage."
Greg Isenberg named a related dynamic that the same economics make more dangerous at the startup level (Novelty-driven AI revenue mimics genuine PMF in the first 60 days and collapses at months 3 to 6). He calls it "vibe revenue": purchases driven by novelty rather than solved problems. Vibe revenue produces the same short-window metrics as genuine PMF: high initial conversion, strong first-month retention, compounding referrals. The signals are statistically indistinguishable in the first 60 days because the underlying behavior is similar, customers repeating a novel experience while it still feels new. The divergence happens between months 3 and 6, when novelty is consumed and no workflow dependency has formed. PMF customers stay because leaving requires changing how they work. Vibe customers stop because nothing holds them. Isenberg's diagnostic: the coming shakeout will be defined by which companies had PMF versus vibe revenue, not by demo quality or early growth velocity.
- Simon Willison, Coding agents collapse all three lifecycle costs of fragile automations, not just the upfront build cost. Coding agents collapse not just creation cost but maintenance and abandonment cost, making previously irrational build decisions rational.
- Greg Isenberg, Novelty-driven AI revenue mimics genuine PMF in the first 60 days and collapses at months 3 to 6. Novelty-driven AI revenue mimics genuine PMF in the first 60 days and collapses at months 3 to 6 when the novelty effect fades.