Architecture breakdown · YAGNI (yagni.app)
Dependable Agent Teams: From Proposal to Merged GitHub PR
Named AI agents (builder, reviewer, PM) that work under human-approved plans with auditable records, sandboxed execution, and realistic evaluation benchmarks.
SYSTEM ARCHITECTURE // YAGNI (yagni.app)
Tbilisi · Remote · NOV 2026 — PRESENT
Dependable Agent Teams: From Proposal to Merged GitHub PR
PROBLEM & CONSTRAINTS
Making AI agent teams dependable in real customer codebases—where every action must be auditable, tenant data strictly isolated, external workspace syncs self-healing, and regressions caught before customers notice.
ARCHITECTURAL SOLUTION
Built tenant isolation and code-review agent routing, made external workspace backfills self-healing, streamlined single-command worktree/CI builds, and built realistic evaluation benchmarks with an early-warning risk radar.
PRODUCTION DATA & CONTROL FLOW:
[Human-Approved Plan] ──► [Named Agent Team: PM · Builder · Reviewer]
[Sandboxed Code Runner + Org Isolation] ──► [Auditable GitHub PR]
[Risk Radar & Stalled-Run Monitor] ──► [Self-Improving Root-Cause Agent]
KEY PRODUCTION OUTCOMES:
- ▸ Tenant isolation: tenant-scoped query guards and sandboxed code execution.
- ▸ Improved code-review agent reliability and evaluation benchmarks: fixed ticket-key collisions across external trackers and built realistic task suites to measure agent changes.
- ▸ Made connector backfills self-healing (stuck-cursor recovery, progress tracking, hard caps) and built an early-warning risk radar for stalled customer runs.
- ▸ Developed an experimental self-improving loop where agents inspect failed runs and propose fixes, and streamlined the monorepo build chain to a single command for worktrees and CI.