Kiro reported on 2 November 2026 that an automated Lean workflow proved 70 properties of models representing parts of its agent harness. Across 26 functions and 96 candidate properties and observations, the team also reported two bounded checks, 18 properties falsified in models and reproduced against analyzed code, and six partial results or observations.[1]

The work covered the shared agent server, Kiro CLI v3 and Kiro Crew. The harness controls context, session persistence and permissions around the language model. Reported theorems include preserving earlier request prefixes during file deduplication and combining parent and sub-agent permission decisions without exceeding the parent’s permission. Queue results assume each store operation is atomic; Lean does not establish that assumption.[1]

The workflow creates a second implementation in Lean, checks it against the original code, then proves statements about that model. Kiro reports 97,938 sampled differential-test cases with no mismatches in the final runs. Those comparisons support correspondence on tested inputs; the universal mathematical results apply to modeled inputs satisfying each theorem’s assumptions. Kiro describes the figures as internal experimental results reviewed by its team.[1]

[1]: Kiro, “Kiro proves properties of its own harness in Lean,” 2 November 2026. https://kiro.dev/blog/harness-properties-in-lean/

Sources

  1. Kiro proves properties of its own harness in Leankiro.dev · 2 Nov 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

25973cf6f396b269aec6e183c456405058838d1f4f07f5ed61416cf39244f8c8