Executive Summary
Three of today's four sources point to the same shift: writing code is no longer the constraint. Specifying behavior and verifying results are. Willison got a serverless acoustic multiplayer game from one Claude Code prompt. Factory, in sponsored content, is selling a "software factory" where engineers write goals and verifiers instead of code. Nate B. Jones argues that Rea, a reverse-engineering tool, matters because vibe-coded apps fail on requirements nobody wrote down. Together they say generation is close to free, and the value now sits in knowing what correct behavior looks like and proving the build matches it. The fourth source, from SemiAnalysis, is on a different layer: ASML is very likely entering die-to-wafer hybrid bonding. That matters for chiplet roadmaps from 2027 onward, but nothing changes for enterprise buyers this quarter.
What Changed
- Prototype cost has collapsed for unusual problems, not just routine ones. Willison's build covers Web Audio signal processing, a modulation scheme, serverless state sync, and game design, from one prompt. He published no latency, error-rate, or device data, so this proves the agent can build it, not that the result works reliably.
- Reverse engineering is now a packaged agent workflow. Rea plugs into Claude Code or Codex and passed 100k GitHub stars in days. It adds no new capability. It removes the skill and setup barrier, which brings the practice to a much wider group of users, including people with no legal training.
- Vendors are pitching orchestration as the place to escape lock-in. Factory sells model-agnostic routing and bring-your-own-model at a $200/month heavy-user tier. The claims are unverified, but the positioning is consistent across the category.
- ASML's bonding entry is probable but unannounced. The evidence is 27 patent families, a supplier report, and its Investor Day roadmap. Its edge is lithography-grade alignment, which goes straight at Besi's logic business at TSMC and Intel. SemiAnalysis expects HBM to skip D2W hybrid bonding until 2029 or later.
Cross-Expert Synthesis
Willison, Jones, and the Factory pitch arrive at the same bottleneck from different directions. Factory says it outright: the engineer's job becomes writing verifiers. Jones gives the reason it is hard. Requirements like Notion's paired clipboard formats or the Windows calculator's context-dependent percent key never appear in a natural-language spec, so an agent building from a spec will produce something that looks right and then breaks. Willison's demo shows the other side of the gap: an artifact that "seems to work," with no measurements behind it.
The disagreement is about where verification comes from. Factory implies the platform supplies it. Jones implies it comes from studying working reference implementations and from domain experts. Willison's practice implies the builder tests directly on real devices. Jones and Willison have the better case. The verifier encodes knowledge about the domain, and a vendor cannot supply your domain knowledge. Treat any orchestration platform that claims to provide verification as providing a harness that you still have to fill.
The SemiAnalysis piece does not join this argument, and forcing a link would be filler. Its value is separate: second-source competition in advanced packaging should slowly lower chiplet cost per die, but throughput of about one wafer per hour and contamination control remain the limits.
Enterprise Implications
Engineering productivity metrics that count output (lines, PRs, velocity) now measure the cheap part. What separates teams is spec quality and test coverage on edge cases, and most organizations do not measure either.
Agent-built prototypes will start showing up as evidence in investment decisions. Without validation under real conditions, leaders will fund ideas that only worked as demos.
Rea creates legal exposure that most enterprises have not scoped. Developers will point it at licensed vendor software from corporate machines. EULA anti-reverse-engineering clauses and derivative-work risk apply whether or not the intent is "learning." The Ledger Reborn case, a reimplementation built from investigating QuickBooks, is exactly the pattern counsel will dislike.
Data-egress policy assumes data moves over networks. Acoustic transport goes around network monitoring, and agents now generate it on request. This is a narrow risk but a real one for regulated environments.
What To Do About It
- Track verification as a separate asset. Fund test suites, eval sets, and acceptance criteria as reusable infrastructure, owned by domain experts, not as a byproduct of delivery.
- Pilot orchestration platforms against your own repos only. Measure merge rate, defect escape rate, and cost per merged change. Require routing logs you can audit for residency and compliance before scaling.
- Issue reverse-engineering guidance now. Define which targets are allowed (open source, your own products) and route licensed commercial software through counsel. Do this before Rea-style tools spread without review.
- Add a validation gate for prototypes. No agent-built demo goes into a funding or architecture decision without measured performance under real conditions.
- Customer question: "When your agent produces something that works, who in your organization knows what it was supposed to do in the edge cases, and where is that written down?" Most cannot answer, and that is the engagement.
Risks and Blind Spots
Three of today's four sources come from one vantage point: individual builders and a vendor. None has enterprise-scale data on defect rates or cost at volume. The Factory material is paid promotion with no evidence behind it. The verification thesis is directionally consistent across sources but unmeasured, so treat it as a strong hypothesis. ASML's timing depends on unresolved engineering problems (feed-forward metrology, die-contact mechanics, bonder-specific contamination) and could easily slip past 2027.