The cloud only moves protocol. Code, secrets and environment stay on your laptop. SubAgents treat top-tier models like rare earth — cheap models do the dirty work in parallel. One request costs cents; an evening's run equals a week of work.
Requirements live in chats, Notion, hallway syncs. Promised yesterday, forgotten today. Acceptance is subjective and untraced. AI helps you write code, but it can't replace the protocol of request → dispatch → execute → review.
Out of vendor subsidy and onto your own dime, the experience cliffs. Claude / GPT bills hit double digits per day. "Good results" and "affordable" are eternally at odds — until you can layer your models.
Code, secrets and SSH keys want to stay local; you also want to dispatch from the subway. SaaS makes you upload code or skip async. What you need is a protocol where the cloud control plane and the local execution plane stay self-consistent.
Real timeline. Each step is tagged with which tool is on stage.
A user dropped a screenshot of a misaligned PC layout on the website. I'm out, on a phone. Open the board in Safari, file a card: title, description, acceptance (with agent-browser self-check), bind to the allwefantasy · auto-coder.chat instance, attach the screenshot. 90 seconds.
Hit Execute. The board packages role + acceptance + instance + attachments into a protocol and pushes it to the auto-coder.chat instance running on my laptop. No IDE, no SSH login.
Local SubAgents in tiers: the high-performance model finds root cause and locates files; cheap models edit and run tests; agent-browser self-checks the CSS. Across 7 minutes the premium model is called only a few times. Bill ≈ a few cents.
The board pinpoints the CSS class. Diff: 12 lines. I glance on the phone, tap Approve → 17:12 it lands in Done. From decision to ready-to-ship: 8 minutes. Never touched a laptop.
Drive home → boot up → pull → reproduce → fix → test → PR → wait for review. An uninterrupted hour is a luxury. The new pipeline shatters that hour into minute-sized fragments — you dispatch on demand, agents work whenever.
Same story, four weapons divided by role
Structured cards + role + acceptance + instance + attachments + verdict + linked session. Every card is a collaboration contract — not Trello-style task management.
Doesn't show up in this story. But WinClaw opens two new channels: (1) check board status / chase reviews from chat; (2) forward screenshots to WinClaw and auto-create cards without ever opening the web.
Doesn't show up here. But when your task becomes "analyze a production metric and file a fix issue," agent_infini and Code Agent share the same pipeline.
Instance registry + heartbeat + remote dispatch. SubAgents tier the models and run worktrees in parallel. /fast edits 8 files in 69.1s for $0.0611 (GPT-5.4 measured).
Multi-agent collaboration is moving from "optional trick" to "default architecture".— subagent-era-is-coming · William