Why we're bullish on loops
The article argues that AI “loops”—self-prompting agents using a goal, context, evaluation, and an agent—outperform single-shot prompts for long-running tasks. It outlines key components, real examples like PR babysitters and bug fixers, and explains why better models, built-in loop commands, and maturing toolchains make loops practical now.
- Opus 4.6 completes half of 12-hour tasks, 6x better than last year's model, and Stripe restructured its whole codebase in a day
- PostHog used a loop to fix a 3-year-old query-engine bug and got an 11% speedup
- Loops work now because of more stable models, built-in loop commands, and mature harnesses/toolchains (subagents, MCP, context compaction)
- The engineer's role shifts from writing code to defining goals, evaluation criteria, and context rather than being replaced
Loops turn agents into self-prompting systems that tackle long tasks without constant human input. You set a clear goal, feed the agent curated context over time (tools, data, even memories), give it ways to evaluate itself (tests, metrics or LLM-based judges), and plug in an LLM agent—Claude Code’s /loop or a custom harness works. Unlike one-shot prompts, loops keep pulling in fresh info and checking progress. You can spin up subagents for parallel work, save tokens, and dodge context overload with compaction and cloud execution.
Real-world loops already boost workflows. A “PR babysitter” loop watches diffs, runs CI tests and keeps hitting “retest” until the build passes. A bug-fix loop reads error traces, writes patches, then runs the suite again. PostHog’s own performance autoresearcher used Karpathy’s loop to squash a three-year-old query-engine bug and speed things up by 11%. Opus 4.6 can now finish half of 12-hour tasks—six times better than last year’s model. Stripe restructured its whole codebase in a day. These wins aren’t hype; they reflect leaps in model stability, built-in loop commands and mature harnesses.
The hype around loops isn’t just vendor marketing. Firms like OpenAI and Anthropic added automations because teams are hitting limits with manual prompting. Subagents isolate sub-tasks. Tools like MCP let loops call external APIs. Context compaction stops windows from filling up. Engineers already run a manual feedback loop: gather analytics, ship tweaks, measure impact, repeat. Loops automate that cycle, handling the small fixes—UX tweaks, flaky tests, bug hunts—so humans focus on strategy, design and empathy.
Critics worry loops will replace coding jobs. In practice, code remains only one piece. Defining goals, crafting evaluation criteria and shaping context demand human judgment. The move isn’t away from engineers; it’s away from constant prompt typing. As these self-driving loops mature, the role shifts toward steering the process rather than writing every line.