Annalea Layton

I architect and direct JUWEL, a persistent AI working environment that turns human intent into completed, verifiable work: chat, memory you control, history and approvals. JUWEL public launch, per the changelog entry of 2026-09-04 (the entry calls it JUWEL OS 1.0). Screens and details →

I architected and directed the security model of JUWEL with a human and agent team: I set the requirements and the test gates, and agents wrote much of the code.

• Approval before acting. 19 tools that act outside JUWEL each need a single-use approval bound to the exact action. An agent cannot approve its own action.
• Signed receipts. Ed25519 signatures, a hash chain, Merkle batches and signed checkpoints, with a public verify page.
• SSRF guard and rate limiting. Web fetch re-checks the resolved address at connect time and on every redirect. Rate limits fail closed, and a checker proves new endpoints cannot skip them.
• Formal methods. TLA+ and Lean models, written by AI under my direction, found a real race in the approval decision path. The fix is specified and not merged yet.

Security and backend details →

Opus orchestrates and verifies, Haiku scouts read, Sonnet workers build one task each, and an Opus closer re-runs the claims and never writes production code. 149-node verification graph, 921 checker scripts. 660 merged pull requests, which I architected and directed, and about half of the repository's non-merge commits (1,222 of 2,460); agents commit alongside me. Swarm details →

• Still Orbit. A Three.js space-survival game. Author-tested local prototype.
• Porcelain Village. JUWEL's world, with 131 author test groups across 18 suites. Work in progress. Game details →