InterWorks: six Titans keep the factory honest overnight
Jun 26: Mnemosyne/Prometheus/Themis/… cron fleet on Curator — review, docs, triage, merge shepherd; deterministic-first Hyperion; shared lessons file; support tickets −50%.
Source: Building a Fleet of Autonomous Agents to Run Your Software's Upkeep
Ben Bausili’s Jun 26 InterWorks note is a production tortoise fleet, not a weekend Claude binge. Curator (their own product) was losing the scheduling fight: docs gaps, flaky tests, PR reviews, and release shepherding all competed with feature work and lost to “later.” Since early 2026, six named agents run around the clock on Claude — review code, write documentation, triage production errors, shepherd releases, watch each other, and open PRs while the team sleeps.
Roles are the reusable design. Mnemosyne (docs) mines assistant chats, tickets, and CI for drift and opens fix PRs. Prometheus (code) takes easy-win issues to PRs and babysits their CI/review. Iapetus (self-healing) reads sibling logs, diagnoses failures, opens fix PRs, stays quiet when healthy. Themis (review) reviews every PR for bugs/security/standards and inherits lessons from the others. Epimetheus (prod triage) ranks Sentry noise by impact and files diagnostic issues. Hyperion (release shepherd) works the weekly merge window — merge approved green PRs in priority order, sync branches, retry flakes, nudge reviewers — and never self-approves. Schedules range from every two hours to weekly so upkeep is continuous, not calendar-dependent.
Build posture is deliberately boring. Most agents are cron Python + Claude Code headless with a three-stage pipeline: Sonnet plans as JSON → Opus applies + lint retry → Sonnet writes the PR comment (cost-split by job type). Epimetheus uses the TypeScript Agent SDK + in-process MCP for Sentry. Hyperion is the thesis punchline: reach for deterministic first — script almost everything, leave Claude one judgment (which approved green PR to merge next). Agents share a lessons file so Themis gets sharper from Prometheus’s mistakes; every script has a failure trap that emails script/line/exit — silent overnight failure is worse than no agents. Each agent is its own GitHub App on ordinary Ubuntu/Python/Node/gh/linters. Claimed receipts: support ticket volume down 50%, faster resolution, developer hours moved to judgment work.
Why it matters for code review
This is continuous harnessed throughput applied to the upkeep queue that usually loses to feature heroics. Review is not a human bottleneck to clear after a burst — Themis never gets “too busy,” and Hyperion’s merge authority is script-bounded so overnight motion cannot auto-approve itself. FOMO that agents aren’t working overnight is answered with a fleet and a self-healer watching the fleet. Pair with Auto mode’s continuity layer and Mouse’s stop-at-merge house rules: always-on only compounds when review, merge, and silence-on-failure are first-class jobs, not afterthoughts.