Alberta says 50 Claude agents audited 466 million lines of code in about 20 hours
The province pegs the equivalent human effort at six and a half years. The numbers come from the vendor and the minister, so read them with that in mind.
- Alberta's Ministry of Technology and Innovation ran about 50 Claude Code agents in parallel to scan 466 million lines of code in roughly 20 hours, per a case study Anthropic published July 6.
- The province estimates a traditional review of the same codebase would take six and a half years.
- Scope: 27 ministries, 1,280 applications, 3,400 repositories, with roughly 95 security controls checked per application pass.
- Red-team agents hunt vulnerabilities, blue-team agents draft patches, and humans review every patch before deployment.
A Canadian province just published the most concrete numbers yet on what agent fleets do to professional services math. According to a case study Anthropic published July 6, Alberta's Ministry of Technology and Innovation deployed roughly 50 Claude Code agents in parallel to scan 466 million lines of code across 27 ministries, 1,280 applications and 3,400 repositories. The pass took about 20 hours. The province's estimate for the same work done by human reviewers: six and a half years.
The setup
Alberta runs the agents in adversarial pairs. Red-team agents probe for vulnerabilities, blue-team agents draft the fixes, and every patch goes through human review before it ships. Each application pass checks roughly 95 security controls. Technology minister Nate Glubish, who wrote up the program in a first-person account, summarized the pitch: "Albertans trust their government with sensitive information, and protection is our responsibility. AI accomplished in hours what traditional approaches would take years."
The program is producing side effects beyond security. One ministry is consolidating 185 legacy applications into 16, and a 25-year-old Java subsidy portal that originally took five months to build was rebuilt in four to five days.
Read the sourcing before the headline
A caveat belongs in the second paragraph, not the footnotes: every figure here originates from Anthropic's case study and the minister's own account. No independent auditor has verified the line counts or the 6.5-year baseline, and vendor case studies exist to sell. The numbers are specific and the program is real, but "Alberta says" is doing work in this headline.
The GaaS math
Even discounted, the shape of the claim is the core agents-as-a-service argument: a security review that would have been a multi-year consulting engagement became roughly a day of metered compute plus human sign-off. That trade, billable years for agent hours, is the same one we tracked when McKinsey turned agents on its own back office. Governments are usually the last buyers to adopt a new procurement model. Alberta suggests the order may flip this time.
Sources: Anthropic case study, Minister Nate Glubish's account, Digital Watch.