Headless Decode on Modal: overnight cron → morning PR inventory

Paul Iusztin’s Sep 10 Decoding AI lesson is the overnight / always-on harness written as something you can deploy, not a vibes demo. After three hours babysitting five local coding agents he lost track of scope; the laptop-lid problem killed anything that needed the machine awake. Lesson 6 of Building a Coding Agent From Scratch takes Decode — the course harness — and makes it remote: drop the TUI for a headless CLI, run the same agent loop on Modal as serverless jobs, and fan out parallel attempts across projects.

Headless mode is decode run "<goal>": same build_agent() loop as the interactive UI, but permissions default to bypass, ask_user is a no-op, stdout keeps only the final answer, and full traces go to Opik. A --max-requests N cap stops unattended doom loops that burn tokens without progress — the failure mode Anthropic’s long-running harness notes also flag. Modal sits between trigger and execution: zero idle cost, fire-and-forget containers. Three entry points hit the same run_task function — decode remote run / attempts from CLI or CI, an authenticated webhook (Slack/Discord/Telegram), and a nightly cron via DECODE_NIGHTLY_CRON. With --sandbox-mode modal each job clones the repo, implements the goal, and pushes a decode/<…> branch (and optionally a PR). decode remote attempts "…" --attempts 3 spins three independent containers so another agent — or you — can compare and forward only the best.

The morning shape is the point. A cron or Linear/Notion webhook can pull tickets tagged for the agent; Modal spawns one container per ticket; you wake to reviewable PRs instead of five terminals. Iusztin cites Ramp’s production pattern: routing queue + session locks, review-ready PRs out, roughly half of merged PRs using that strategy. Orchestrator-plus-GPU-sandbox subagents extend the same idea for parallel fine-tunes. Serverless fits bursty overnight work (scale to zero when idle); always-on interactive TUIs may want different infrastructure. Cost still needs a ceiling — commenters flag that five agents × N attempts × nightly cron scales with ambition, not merged output, unless you name a monthly cap or cost-per-merged-PR.

Why it matters for code review

Overnight agent motion only compounds when delivery stops at a reviewable inventory with independent traces — not when five local sessions demand your presence until the lid closes. Pair remote headless triggers (cron, webhook, CI) with attempt isolation and a morning pickup habit: continuous throughput without that boundary just relocates babysitting into the review queue. Pair with verify-first rejection-tax receipts and queue design for agent volume: remote harnesses mint PRs at machine pace; review capacity and live verification still decide what ships.