V5 scanned its own codebase in 35.2 seconds and returned 406 findings. Seventy-five of them were security issues — hardcoded secrets, missing input validation, over-permissive CORS, the standard patterns. The grade at the bottom of the report was a D.
V5 is my AI code quality scanner. The codebase it scanned was its own, and its own codebase was written with best-available AI tools by a security-conscious practitioner who has shipped production software for years. The tool was scanning the work of a careful author, and it came back with 406 things to fix.
That report is the whole argument for PRED-007. Not the part people usually quote. The other part. The part that makes the 10x cost drop either true or a lie, depending on whether you budget for the D.
PRED-007 — By December 2028, agent-first companies — where agents handle 60%+ of code generation, testing, deployment, and monitoring — will spend 10x less per shipped feature than traditional human engineering teams. Measurable across developer salary equivalents, time-to-ship, and total cost of delivery.
Confidence: 4 out of 5.
The cost trajectory is not controversial. GitHub Copilot users complete tasks 55% faster in co-pilot mode, and that is the human-plus-AI configuration, not autonomous agents. API inference costs dropped roughly 10x between 2023 and 2025 — GPT-4 launched at $30 per million input tokens in March 2023; equivalent capability costs $3 or less in 2026. Agent delivery of whole features, not just suggestions, plausibly compounds those gains to 5–10x end-to-end productivity. My own ventures — V4 security scanning, V5 code quality — exist because that math actually works at the scale one person can afford.
Cloud SaaS did the same thing to on-premise software between 2010 and 2020, cutting delivered cost by 60–80% in CRM, collaboration, and HR. Agent-first delivery is the next structural shift, and the mechanism is more radical: salary cost is typically 55–67% of operating expenses, and agent-first companies compress that line toward API inference, which is currently a fraction of the equivalent labour cost.
That is the pro-case. Now the part the chapter does not say explicitly, which is implicit in PRED-013 (the AI-generated code breaches prediction) and which the V5 self-scan makes unavoidable.
The 10x is conditional. The condition is a quality tax.
The 10x number is clean only if AI-specific quality infrastructure exists at the same scale as code generation. Without it, the 10x is a lie — you are not cutting cost, you are moving cost from the salary line to the breach-remediation line. The cost still shows up. It just shows up later, on a different budget, and often in the CFO’s Q4 surprise column.
The quality tax is made of concrete line items. I name them because a tax you cannot itemise is a tax you cannot budget.
AI-specific SAST. Not generic static analysis. Scanners that understand the patterns AI models actually generate — hardcoded secrets, phantom dependencies, over-permissive CORS, unsafe deserialisation.
Phantom-dependency detection. Catching the imports an AI invented that do not exist on npm, or do exist and belong to someone else.
Hardcoded-secret scanning. At generation time, not at deploy time. V5 sees this pattern more than any other.
Agent-PR review discipline. The human review seat, with receipts. Not rubber-stamping. Actual reverts when the agent is wrong.
Budget those four lines at 15–25% of the “saved” engineering cost and PRED-007 holds. Skip them and you ship faster for two years, then take a breach or a quality reputation loss that wipes out three years of savings in a quarter.
So the prediction is not “ship features at 1/10th the cost.” The prediction, written honestly, is “ship features at 1/10th the cost if you pay the quality tax, and take a catastrophic loss on a two-to-three-year delay if you don’t.” The companies that will make PRED-007 true are the ones who put the quality tax on the budget before they start celebrating the salary line. The companies that will make PRED-013 true are the ones who do the opposite. Same prediction seen from two sides of the same spreadsheet.
The published falsification trigger:
If by December 2028, no credible study shows more than 3x cost reduction for agent-first delivery, or if quality gaps make the comparison invalid, this prediction is wrong.
That is the stake as written. Here is the way I am most likely to lose it.
The second clause. Not the first. The 3x number will be cleared — agent-first shops are already cruising past it in narrow measurements. The quality gaps clause is where the prediction actually lives or dies. If by 2028 the credible studies of agent-first cost savings start carrying asterisks large enough to cover the claim — “before defect remediation”, “excluding security incidents”, “pre-revert” — then the 10x is technically reported and substantively invalid. That is the failure mode. And I am watching for it in every vendor-published productivity study coming out of the big model providers right now.
I want your cost breakdown with the quality tax included.
If you work inside an agent-first shop — or a shop that is 60%+ agent-generated on its code path — send me your feature-delivery cost breakdown for any 2025 or Q1 2026 feature, including post-ship defect remediation costs (reverts, patches, incidents, customer-reported bugs attributable to the delivery). Anonymise whatever you need to.
I will publish the distribution of net cost after quality tax, versus traditional-team baselines, at atin-agarwal.com/predictions/pred-007-saas-feature-cost-drop/, quarterly. If the median net cost across respondents is within 3x of traditional teams, PRED-007 in its stated form is wrong — and I want to know that before 2028, not after.
Counter-data welcome. A single well-measured counter-example is more valuable to me than ten self-reporting success stories.
If you are a founder scaling an agent-first shop: budget the quality tax at 15–25% of your “saved” engineering cost before you declare a cost win on the board deck. If your model assumes a 10x win net of quality spend, run the arithmetic one more time with the tax included. The win is still large. It is not 10x.
If you are a CFO: your “feature unit cost” metric is invalid unless it includes post-ship defect remediation. Make the finance team rebuild the line. The version without it is going to look beautiful until the first incident, and then the number you wish you had been tracking is the one the board will ask for.
If you are an engineer in an agent-first shop: your 2028 seat is in the quality-tax collection department. Start there on purpose. “I reviewed 4,200 agent-authored PRs last quarter and reverted 6%” is the résumé line, and the career path. The people who build the quality tax infrastructure in 2026–2027 own the category by 2029.
If you are a buyer evaluating an agent-first SaaS vendor’s pricing: ask the vendor what percentage of their cost base is quality infrastructure. Any answer under 10% is a vendor who has not yet paid the tax. Their price is going up or their quality is going down — pick one.
This issue is drawn from Chapter 9 of The AI Agent Economy — 15 falsifiable predictions with dates, numbers, and explicit triggers for being proven wrong. Pre-order on Kindle — $9.99. Release July 1, 2026. atin-agarwal.com/books
Read the full PRED-007 entry on the public tracking page → atin-agarwal.com/predictions/pred-007-saas-feature-cost-drop/
Previous issue: Issue 06 — 1 in 5 Series A rounds will go to teams smaller than this email thread Next issue: Issue 08 — Three SaaS categories will lose half their price by 2029 — and one isn’t on the book’s list
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.