Claude Loops¶
Two disciplines, kept separate.
This knowledge base covers two related but distinct things. Most confusion about working with coding agents comes from treating them as one.
| Part I — Loop Engineering | Part II — Developing with Claude Code | |
|---|---|---|
| What it is | Designing a system that prompts an agent for you — it fires on a trigger, works, verifies, and stops | You and Claude Code building software together, iteratively, with you in the loop |
| You are | Not at the keyboard | At the keyboard |
| Scope | General, tool-agnostic | Claude Code specific, version-stamped |
| Start at | The Loop Contract | Choosing Your Mode |
If you are not sure which you need, start at Choosing Your Mode. It is the router: three tests a task must pass before a loop is worth building, and a decision table for everything that fails them.
Which one is the default?¶
For building software, the honest answer is Part II.
"most effective coding agent use is a complex, highly iterative process"
— Andrew Ng, …Using Coding Agents, 2026-09-04 — explicitly not autonomous long-horizon execution
"Most coding tasks involve fewer truly parallelizable tasks than research, and LLM agents are not yet great at coordinating and delegating to other agents in real time."
— Anthropic, How we built our multi-agent research system, 2025-06-13
Part I is not the lesser half. It is where recurring, well-specified, verifiable work belongs — triage, sweeps, dependency bumps, docs sync, digests. That work is genuinely better done by a loop than by you, and the payoff compounds. But it is a class of work, not the default for everything.
Part I — Loop Engineering¶
General and tool-agnostic. Claude Code appears here as an example, never as a prerequisite.
The central act is designing the loop — deciding what it is for, how it does it, when it fires, how much it may spend, and how you know it's done. That last question is the one most loops get wrong. The Loop Contract is the instrument for answering all five.
1. Foundations¶
- 1.1 The Paradigm Shift — why a correction loop beats a single turn
- 1.2 The Core Agent Loop Cycle
- 1.3 Loop Maturity Model
- 1.4 Context vs. Loop Engineering
- 1.5 The Factory Model
2. Designing a Loop¶
- 2.1 The Loop Contract — the design spine
- 2.2 Goal Engineering
- 2.3 Harness Patterns
- 2.4 Loop Patterns Catalog
3. Verification & Failure¶
- 3.1 Verification — the non-negotiable foundation
- 3.2 Common Failure Patterns
- 3.3 Human-in-the-Loop Escalation
4. Scaling¶
Part II — Developing with Claude Code¶
Concrete and version-stamped. Platform facts here carry the version they were true in, because
they move — this knowledge base had to correct eight stale facts in v3.0.0 for exactly that reason.
5. Start Here¶
- 5.1 Choosing Your Mode — the router
6. The Workflow¶
- 6.1 The Development Workflow — plan → execute → operate, and the five skills
- 6.2 Session Architecture — one session or many
- 6.3 Explore → Plan → Implement → Commit — plan mode's keybindings, what it does and does not auto-approve, and the read-only subagents that do the exploring
7. Your Setup¶
- 7.1 CLAUDE.md
- 7.2 Skills
- 7.3 Hooks
- 7.4 Permissions & Auto Mode
- 7.5 The Six Building Blocks
- 7.6 Memory Patterns
- 7.7 Context Management
8. Parallel Work¶
9. Running Unattended¶
10. Cost & Safety¶
11. Reference¶
Stay current¶
A daily loop scans the community for new practice and folds verified findings back into these docs. It is itself a Part I loop — recurring, well-specified, and verified — which is why it exists.