Theo explains how he combines coding subscriptions, a shared proxy, remote computers, and parallel agent tasks to reduce the attention needed to ship software. The walkthrough covers subscription economics, reset timing, verification, T3 Code's task inbox, exploratory delegation, and overnight work. It is an opinion-led technical walkthrough with demonstrations, rather than a complete installation tutorial.
Video: Theo - t3․gg on YouTube, 1:06:36. Source: verbatim transcript. Prices, model names, product behavior, and risk assessments are recorded as presented in the video, not verified current guidance.
Topic breakdown
- Agent maxing: purpose and limits, 0:00–1:30 and 2:43–5:23.
- Sponsor: Parallel search tools, 1:30–2:43.
- Subscription economics and API boundaries, 5:23–12:41, with qualifications at 15:38 and 19:43.
- Proxy access and fleet management, 12:41–18:04 and 22:12–26:18.
- Routing, caches, and reset timing, 18:04–22:12 and 26:18–29:09.
- Verification and problem framing, 29:09–38:35.
- T3 Code: manage threads as a task inbox, 38:35–50:07.
- Experiments and work triage, 50:07–59:05.
- Overnight work and hardware choices, 59:05–1:03:10.
- CI bottlenecks and the human role, 1:03:10–1:06:36.
Key takeaways
- Theo estimates $200 subscriptions provide about $8,000 of Claude API-equivalent usage, including $4,000 of Fable, or $12,000 of Codex usage. These are usage valuations, not cash expenditure or guaranteed allowances.
6:116:51 - His Opus 5.5 correction weakens the case for multiple accounts: he finds one subscription hard to exhaust and says a proxy may no longer be necessary.
19:4320:05 - Account affinity preserves account-specific caches; reset-aware routing favors allowance that would otherwise expire first.
18:3420:5121:19 - Reset clocks differ in his account: Codex resets restart the weekly timer, while Claude resets refill capacity without moving the renewal date.
26:1827:3528:02 - He estimates 85–90% of his token usage goes to verification. His recommended workflow checks code before merge and makes reverting a bad change take less than 15 seconds.
30:4831:17 - Give agents the problem early and request the whole deliverable, including tests, a reviewable PR, a test environment, and monitoring, to reduce repeated interruptions.
34:2946:2146:4147:02 - His overnight example includes over 12 concurrent tasks lasting from about 30 minutes to 18 hours. Remote execution lets those tasks continue after the laptop closes.
58:0158:1859:05 - Heavy CI can be the bottleneck even when orchestration uses little CPU. Remote CI or queued builds can preserve local capacity.
1:01:381:03:101:03:33 - He reserves APIs for public user-facing requests and explicitly acknowledges possible bans or terms violations for the proxy arrangement.
12:0814:41