Inside Anthropic’s Loop Engineering: Designing Self‑Running Agent Systems

What Is Loop Engineering?

In June 2026 three senior engineers—Peter Steinberger, Boris Cherny, and Addy Osmani—converged on the idea of replacing manual prompts with a system that runs agents autonomously. Loop Engineering is defined as replacing the human who gives the agent instructions with a designed system that does so automatically.

Four‑Layer Stack

The methodology adds a fourth layer on top of the existing three: Prompt Engineering (write a single prompt), Context Engineering (select what to retrieve and summarize), Harness Engineering (arm a single run with tools), and Loop Engineering (schedule the harness to run repeatedly).

Five‑Step Loop and Six Components

The loop consists of Discovery → Handoff → Verification → Persistence → Scheduling . Each step is realized by concrete components:

Automations : timers or triggers that implement Scheduling.

Worktrees : isolated directories for parallel agents, used in Handoff.

Skills : permanent project knowledge (e.g., SKILL.md) used in Discovery.

Connectors : MCP‑based hooks to external systems, supporting Persistence and Discovery.

Sub‑agents : separate agents for code generation and code review, enabling Verification.

Memory : disk‑based persistent state for Persistence.

Generator / Evaluator Separation

Anthropic engineers observed that letting the same agent both generate code and judge its quality leads to self‑approval bias. Calibrating a dedicated Evaluator agent that defaults to skepticism, performs action‑based checks, and hands off final decisions to a smaller model proved more reliable.

Failure Modes

Skipping any of the five actions creates a predictable anti‑pattern:

Verification – “Nodding Loop”: never says “no”, approves bad code.

Persistence – “Amnesiac Loop”: loses state each day.

Scheduling – “Manual Loop”: runs only for a demo.

Discovery – “Blind Loop”: humans still decide what to do.

Handoff – “Tangled Loop”: parallel agents clash on the same directory.

Real‑World Cases

Osmani’s Morning Loop reads yesterday’s CI failures, opens isolated worktrees, drafts fixes with a sub‑agent, reviews with another, auto‑submits PRs, and records state for the next day.

Stripe’s Minions merges over 1,300 machine‑written PRs per week without any human‑written code. The pipeline interleaves deterministic orchestration (hard‑coded context gathering, lint gates) with LLM‑generated code, ending with human review. Stripe’s core claim: reliability comes from constraint quality, not model size.

Scheduling Options compare Cloud (hourly, no local access), Desktop (minute, local file access), and /loop (local session, minute, local file access).

Hidden Costs

Four silently accumulating costs are identified:

Verification Debt : untested outputs pile up until a release day explosion.

Comprehension Rot : developers fall behind understanding code written by loops.

Cognitive Surrender : humans stop reviewing because loops appear reliable.

Token Blowout : spawned helpers, retries, or infinite loops can burn entire budgets.

Safety Discipline for the First Loop

Read a Sample, Always : daily read a representative output and explain its behavior.

Cap Before You Ship : set per‑run, daily, and retry limits to bound risk.

Keep One Door Open : always retain a human checkpoint to intervene if needed.

Final Formula

Stop giving agents direct instructions; instead, design the system that gives those instructions, approaching it as an engineer rather than a button‑pusher.

https://x.com/0xCodez/status/2069736449902027136
https://drive.google.com/file/d/1qzKI4DKnyHRpXK1J3ATPqwaqLc0iNu-M/view