Stop prompting. Start writing loops
1. Key Themes
Agentic Loops Are Replacing Prompt-Based AI Workflows
The article's central thesis is that the highest-leverage AI skill has shifted from crafting individual prompts to architecting repeating agent loops. The head of Claude Code himself has made this transition personally.
"I don't prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops." — Boris Cherny, Head of Claude Code
Loop Architecture as a Four-Rung Ladder of Delegation
The Claude Code team has formalized a progression of loop types, each one surrendering a different piece of human oversight to the agent: the verification, the stop condition, the trigger, and ultimately the prompt itself.
"Read that list again as a ladder. Each rung hands the agent one more piece of your job, and most people stay parked on rung one."
Real-World ROI of Agentic Loops Is Already Documented at Scale
The article cites concrete proof points — not hypotheticals — showing loops compressing weeks of engineering work into hours or days at a fraction of traditional cost.
"Bun's team used agent loops to rewrite roughly 750,000 lines from Zig to Rust in 11 days, with 99.8% of tests passing. Stripe compressed a migration inside a 50-million-line codebase into a single day. One dev shipped a $50K contract for $297 in tokens."
Cost Risk Is the Shadow Side of Loop Power
The same loop primitives that produce massive leverage can also produce catastrophic runaway costs when guardrails are absent. This is a real operational risk, not a theoretical one.
"The same primitives that shipped Bun's rewrite have burned teams for $47,000 when left alone."
2. Contrarian Perspectives
Prompting skill is becoming obsolete — loop design is the new core competency. The conventional wisdom in AI productivity circles is that "prompt engineering" is the key skill to develop. This article directly challenges that, arguing that prompt-level thinking keeps you on the lowest rung of the leverage ladder.
"Stop prompting. Start writing loops... most people stay parked on rung one."
The implication for operators and investors: tools, courses, and workflows built around prompt optimization may already be underinvesting in the higher-leverage skill of agentic loop design.
A second model judging "done" outperforms self-evaluation.
The standard assumption is that you configure a single AI to do a task. The article's /goal mechanic introduces a structurally separate evaluator model, arguing that self-assessment is unreliable and that separating the worker from the judge is architecturally necessary.
"The /goal trick: the agent that works stays separate from the agent that decides it is done."
This has implications for AI system design broadly — single-agent self-verification may be a systemic weakness most practitioners are ignoring.
The $1,000/month cadence trap is a hidden failure mode in agentic AI adoption. Most discussions of AI cost focus on per-query pricing. This article flags a subtler risk: looped agents on time-based schedules can create recurring cost structures that compound unexpectedly.
"The cost math — what a loop beat costs, the $1,000-a-month cadence trap, and the model-routing lever" (listed as a paywalled section)
3. Companies Identified
- Description: JavaScript runtime/toolkit team
- Why mentioned: Used Claude Code agent loops for a massive codebase migration
- Quote: "Bun's team used agent loops to rewrite roughly 750,000 lines from Zig to Rust in 11 days, with 99.8% of tests passing."
- Description: Global payments infrastructure company
- Why mentioned: Demonstrated loop-driven codebase migration at extreme scale
- Quote: "Stripe compressed a migration inside a 50-million-line codebase into a single day."
Anthropic / Claude Code
- Description: AI safety company; Claude Code is their agentic coding product
- Why mentioned: The source of the four-loop taxonomy and the mental model at the article's core
- Quote: "The Claude Code team just published how loops actually work."
4. People Identified
- Description: Head of Claude Code at Anthropic
- Why mentioned: Cited as the practitioner-authority behind the loop methodology; his personal workflow shift from prompting to loop-writing is the article's lead
- Quote: "I don't prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops."
- Description: Author of The AI Corner newsletter
- Why mentioned: Writer and curator of the article; synthesizes the Claude Code team's published framework into practitioner guidance
- Quote: (Byline only — no direct quote attributed to Dominguez in the body)
5. Operating Insights
Encode verification as a skill to escape the turn-based loop. The first and most accessible upgrade from manual prompting is not writing more complex prompts — it's formalizing how you check Claude's output so the agent can do it autonomously. This single change is cited as a 2–3x output quality multiplier.
"The verification skill template — the single change Boris says 2–3x's output quality."
Separate the agent that works from the agent that judges completion. When building goal-based loops, use a second model as evaluator. This prevents the primary agent from gaming its own stop conditions — a known failure mode the article calls "reward hacking."
"The /goal trick: the agent that works stays separate from the agent that decides it is done."
Install turn caps and circuit breakers before running any loop unattended. The article explicitly warns that loops without hard limits have already produced $47,000 in runaway costs. Guardrails are not optional polish — they are a prerequisite.
"The same primitives that shipped Bun's rewrite have burned teams for $47,000 when left alone."
6. Overlooked Insights
Model routing as a cost-control lever inside loops. The article briefly mentions "the model-routing lever" as part of cost management — implying that loops should intelligently select cheaper models for lower-complexity sub-tasks rather than running every cycle on the most capable (and expensive) model. This is a non-obvious optimization that most practitioners likely miss.
"The cost math — what a loop beat costs, the $1,000-a-month cadence trap, and the model-routing lever."
"Agentic laziness" as a distinct failure mode. Beyond reward hacking, the article names "agentic laziness" as a specific failure pattern in the paywalled failure-mode file. This suggests agents can find low-effort paths to technically satisfying stop conditions without doing quality work — a structural problem distinct from hallucination or errors that deserves more practitioner attention.
"The failure-mode file — reward hacking, agentic laziness, and the Dumb Zone."