Theo argues agent loops should replace manual prompt-by-prompt coding
Theo (t3.gg) shares his shift from manually shepherding coding agents step-by-step to building automated loops where agents prompt, review, and re-trigger each other.
Score breakdown
The video documents a practitioner's firsthand shift from manual agent orchestration to fully automated agent loops, illustrating a concrete change in how agentic coding workflows are structured in practice.
- 01Theo argues developers should design loops that prompt agents, rather than writing prompts themselves.
- 02He previously found loop-based approaches like 'the Ralph loop' increased agent error rates and felt unproductive.
- 03His prior workflow was manual: plan review, step-by-step direction, agent review, and feedback handoff — all orchestrated by him personally.
Theo opens with the framing that developers should no longer be writing prompts for coding agents — they should be designing loops that do the prompting. He admits this idea didn't land with him initially: while he had seen loop-based approaches before (he mentions "the Ralph loop" by name), he found they increased the error rate of agent-generated changes and didn't feel productive in practice. His default workflow remained manual orchestration — asking a model to make a plan, reviewing it, directing each step, having a second agent review the output, and feeding that feedback back to the first agent. He was the one running the loop by hand.
He credits a figure he calls "Pete" as someone who was ahead of the curve on agent-loop thinking.
His perspective shifted after he began building automated systems himself: agents that review code and trigger re-reviews, systems that watch pull requests and issues on external repos for updates, and a "Hermes agent" that brings context to him proactively. He credits a figure he calls "Pete" as someone who was ahead of the curve on agent-loop thinking. After shipping code with these loop-based systems, Theo says he now believes most agent runs should not be initiated by a prompt the developer personally wrote — a position he describes as something he never expected to hold. The video is truncated before he fully elaborates on his findings, and includes a sponsored segment for Magic Patterns, an AI-assisted front-end design tool.
Key facts
- 01Theo argues developers should design loops that prompt agents, rather than writing prompts themselves.
- 02He previously found loop-based approaches like 'the Ralph loop' increased agent error rates and felt unproductive.
- 03His prior workflow was manual: plan review, step-by-step direction, agent review, and feedback handoff — all orchestrated by him personally.
- 04He shifted after building systems where agents review code, trigger re-reviews, and watch pull requests and issues on external repos.
- 05He references a 'Hermes agent' that proactively brings context to him rather than requiring him to go find it.
- 06He now believes the majority of agent runs probably should not use prompts the developer wrote directly.
- 07The video is sponsored by Magic Patterns, an AI front-end design tool.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 19, 2026 · 10:25 UTC. How this works →