LoomStack
CONCEPT

Loop Engineering

You stop prompting agents. You design systems that prompt them. Loop engineering is the next abstraction layer above prompt engineering.

4/4 steps100% automated1 rework loop
Code Agent
sonnet-4· 3m
Test Agent
sonnet-4· 7m2/8 FAIL
2 tests failing — retry with fix
Code Agentretry 2
sonnet-4· 5m

Fixed edge case in retry handler

Test Agent
sonnet-4· 7m8/8 PASS

Definition

What is loop engineering?

The shift from prompting to systems design.

Boris Cherny, creator of Claude Code at Anthropic: “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.”

Loop engineering is designing a small system that finds work, hands it to agents, checks the output, records what's done, and decides what's next. You define the goal and the constraints. The system pokes the agents instead of you.

The progression: autocomplete (2023) → prompting AI to write code (2024) → running multiple agents in parallel (2025) → writing loops that do the prompting themselves (2026).

“You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents.”

Peter Steinberger, OpenAI

Primitives

The five building blocks

Both Claude Code and Codex ship all five. A loop needs primitives and a place to persist state.

01

Automations

Discovery and triage on a schedule. Define a task, a cadence, and a stopping condition. The loop finds work and acts on it while you sleep.

/loop, /goal, hooks, GitHub Actions

02

Worktrees

Isolated branches so parallel agents don't collide. Each loop gets its own checkout. One agent's edits can't touch another's working directory.

git worktree, --worktree flag

03

Skills

Codified project knowledge so agents stop re-learning every session. Conventions, build steps, architecture decisions written once and read every run.

SKILL.md files, $skill invocation

04

Connectors

MCP-based integrations that let the loop touch your real systems. Issue trackers, databases, staging APIs, Slack.

MCP servers, plugins, tool permissions

05

Sub-agents

Split the maker from the checker. The model that wrote the code doesn't grade its own homework.

.codex/agents/, .claude/agents/

+

State / Memory

A markdown file, Linear board, or any store that outlives the conversation. The agent forgets between runs. The repo doesn't.

Example

What a loop looks like

Daily triage loop — concrete, not abstract.

6:00am

Automation triggers. Calls a triage skill that reads CI failures, open issues, and recent commits.

6:01am

Findings written to state file. For each actionable item, opens an isolated worktree.

6:02am

Sub-agent drafts the fix. Second sub-agent reviews against project skills and tests.

6:08am

Connectors open PR, link Linear ticket, ping Slack. Anything it can't handle lands in your triage inbox.

Next day

Loop picks up where yesterday stopped. State file remembers what was tried, passed, remains open.

You designed it once. The loop finds work, acts on it, verifies itself, and reports back.

Team scale

Where loops break at team scale

Individual loops work. Organizational loops need coordination.

Developer A's morning triage loop “fixes” a test that Developer B's refactoring loop intentionally changed. B's loop reverts A's fix that night. Neither person is awake for either event.

Problem 01

No shared context

Each developer's loops operate in isolation. No loop knows what other loops are currently changing across the org.

Problem 02

No governance

A loop can violate architectural boundaries, touch services it shouldn't, or deploy at 3am.

Problem 03

No observability

When ten loops run across a team, nobody has a unified view of what they're collectively doing to the codebase.

Problem 04

No coordination

Worktrees isolate one developer's loops from each other. They don't isolate one developer's loops from another's.

Platform

LoomStack: the orchestration layer for loops

Individual loops sit above individual prompts. LoomStack sits above individual loops.

Gap 01

Shared context

Every loop reads from and writes to shared organizational memory. Loops know what other loops are doing.

Gap 02

Policy enforcement

Governance rules evaluated before a loop's changes land. Loops respect boundaries without each developer configuring them.

Gap 03

Observability

Unified view of all loop activity across the org. Trace any production change back to the loop that made it.

Gap 04

Coordination

Conflict detection across loops before they compound. Developer A's loop and Developer B's loop don't work against each other.

Coordinate your loops at organizational scale

LoomStack is the orchestration layer above individual loops. Shared context, governance, observability, and coordination so your team's loops work together.