Executive Summary
Generating software is no longer the bottleneck. Verifying it, specifying it, and being accountable for it are. All four sources point the same way from different angles. Latent Space shows a $43K/yr SaaS tool can be cloned for about $11K, but only after heavy specification work and testing by real users, because the AI judge picked the wrong winners. Jones argues that verifiable rewards produced superhuman math but no superhuman engineer, because the hard part is knowing which question to ask. Huntley shows that rules written as prose drift away from what agents actually do, while deterministic gates hold. Meta's Muse release adds a fifth frontier vendor, which makes multi-model architectures more valuable and vendor benchmarks less trustworthy. The enterprise money is now in the verification layer, not the generation layer.
What Changed
- Mid-tier SaaS is now a build candidate. Latent Space received 72 submissions, and 5 of them were good enough to adopt. Their estimate for a clone is about $1K in tokens plus $10K of expertise. In return you own the maintenance.
- AI judges have failed a public field test. The top seven apps scored within 2.3 points of each other. Six of the judge's top 10 were rated weak by the people who would use the tool. The real winner ranked 19th.
- Mixing models is now standard practice. 40 of 66 teams used both Claude Code and Codex, and every finalist used GPT. The rule they followed was that no model approves its own work.
- Meta is at the frontier, but it is expensive and short on capacity. Muse Wonderful scores 56 on Artificial Analysis, against 58 for Opus 5.5. It costs $4.52 per task, about 6x GPT-6.1 Soul Max. Users get 429 rate-limit errors even on the $200 tier, and the strongest variant is locked inside a coding subscription.
Cross-Expert Synthesis
The sources agree on what limits progress, and they partly disagree on what that means for build decisions.
Agreement: verification sets the limit. Jones's revoked-permission example (tests pass, a stale index still leaks content) is the same failure Latent Space found in practice: a headshot upload the public page never read, and duplicated features that weren't connected to each other. Both say agents pass the checks they are given and miss the checks nobody thought to write. Huntley's hook is the practical answer to one part of this. Once a team knows the question, it should be encoded as a deterministic gate so it doesn't depend on a context window that degrades over a long session. Cross-model checking addresses another part, since agents are lenient toward their own output. Neither one solves Jones's underlying point: finding the question in the first place still takes a human who is in contact with the domain.
Tension: is cheap building a gift or a burden? Latent Space treats ownership as the payoff, because domain experts file bugs and fixes come back within minutes. Jones warns that faster generation just piles more review onto a small group of senior engineers. Both can be right, because they describe different conditions. The SaaS replacement worked because the acceptance testers were the end users and the domain was bounded. Event management has low security exposure and failures are visible in the UI. Move the same approach to a system with access control, compliance, or data leakage risk, and Jones's critique applies in full.
Eval distrust applies to vendors too. Meta's charts left out Grok 4.7 on the legal benchmark, and Muse led on coding benchmarks while trailing in hands-on generation tests. Whether it is an AI judge scoring apps or a vendor scoring itself, the lesson is the same: only evaluations on your own workloads, judged by your own users, count.
Enterprise Implications
The control surface for agentic development is the existing delivery pipeline: hooks, linters, CI gates, fitness tests. AGENTS.md files and prompt libraries are not that surface. Organizations with mature pipelines can absorb agent output. Those without them will see output quality vary and code drift from internal standards, and senior reviewers will burn out.
Reviewing the SaaS portfolio is now a real infrastructure question. The target is poorly loved mid-tier tools: $20K to $100K a year, low regulatory exposure, slow vendor change requests, and most of the bundle unused. Replacement moves the 2 a.m. pager from the vendor's team to yours. That trade only makes sense where internal ownership produces improvements the vendor would never ship.
Being able to route across vendors is now a hard requirement, both for leverage and because capacity varies between providers. Muse is a candidate for evaluation, not something to migrate to.
What To Do About It
- Audit guardrails before scaling agent seats. List which engineering standards are enforced mechanically and which exist only in documents. Convert the top ten policies (telemetry libraries, banned dependencies, secrets handling, audit logging) into checks that run in CI, not just local pre-commit hooks, which agents can skip with
--no-verify. - Run one SaaS replacement pilot. Choose a bounded tool with low exposure. Fund a small engineering team paired with domain owners who act as testers. Use automated evals for hill-climbing only, and let real users decide acceptance.
- Require cross-model verification. Separate the planner, the implementer, and the reviewer across model families. Write it into the agent workflow standard.
- Measure review capacity, not generation throughput. Track senior reviewer load and the age of the review queue. If generation speeds up and those numbers get worse, the program is creating debt.
- Customer question: "If an employee's access is revoked, which of your AI systems would still show them what they used to see, and who would know to test for that?"
Risks and Blind Spots
The Latent Space numbers come from one low-stakes domain with a motivated competitor pool and a $10K prize. Many enterprises don't have $10K worth of expertise sitting idle, and the replacement cost leaves out long-term maintenance, security review, and the risk of losing the engineers who built it. Huntley's result rests on a single anecdote and covers only violations that can be detected statically in a diff. Jones offers no data, and his transfer-gap claim could fail quickly if labs find ways to train verification of the kind his example needs. Meta's unprompted "lineage" output is minor, but it is the kind of anomaly that belongs in red-team suites before any production use. Finally, all four sources assume that domain experts have the time to act as testers. In most enterprises they don't, and that is the constraint nobody here priced.