Lore

Loop Engineering

A meta-level practice of engineering the loop or goal that keeps re-prompting a coding agent, so it keeps working without the human retyping prompts after every step. Sits above Context Engineering in the AI-Native Developer Workflow.

Mechanism

Instead of issuing one prompt per task, the human sets a standing 'goal' (e.g. "groom all the MVP issues") or a recurring loop prompt, and the agent self-continues — pulling the next item and repeating the same action — until the objective is met, rather than stopping after a single item.

Remote Unsupervised Execution

Two pieces of infrastructure make it practical to let an agent run a loop unattended for long stretches:

Autonomous Continuation Mechanisms

Two built-in mechanisms implement loop engineering without a manual restart:

Claude Code's 'Loop' Feature

Claude Code's 'Loop' Feature

Claude Code ships a native Loop feature that re-sends a scheduled prompt (e.g. every 30 minutes) to the same session, letting the agent check in on long-running work without a human re-prompting it. The loop can be told to stop once a job is judged complete.

This is Claude Code–specific — engines without native loop support (per the presenter, this includes Codex) don't get it for free. The documented workaround is to replicate the same effect via stop hooks or by periodically sending messages into a tmux session the agent is attached to, keeping the loop-engineering pattern usable across tools even when the underlying primitive differs. See AI Dev Tool Taxonomy for other cross-tool capability gaps.

Why Non-Engineers Must Learn It by Building

PMs and designers need to deeply understand how agent loops work, not just conceptually — agent-coding harnesses are treated as analogous to "the operating systems of the '90s": foundational infrastructure that shapes everything built on top, and that only becomes intuitive through direct use. The recommended path is to have non-engineers build small things directly with agent harnesses, internalizing the loop's behavior firsthand rather than learning it only in the abstract — consistent with the broader claim that Agency as the AI-Era Differentiator is built by tinkering in code as a malleable material, per Malleable Software.

Agent as LLM-in-a-Loop (Linear)

Jacob Shumway (Linear) defines an agent, at the engineering level, as an LLM called repeatedly in a loop rather than invoked once for a single prompt/response — the loop is what lets it reason toward a goal instead of following a scripted instruction. Linear's team separates this from the product definition of "agent": a packaged subsystem presented to users behind a single facade (e.g., "the Linear agent") that may itself be built from several such loops, tools, and guardrails. Building a production agent means solving both definitions at once — the reasoning loop, and the product packaging around it.

The operational pattern: define a goal or milestone, give the agent tools to build its own context, then let it run question → tool-call → answer turns until it has gathered enough information to synthesize a final response (e.g., write and file an issue). See Context Engineering for why Linear favors this over pre-loading a scripted prompt.