What Are AI Agent Constructs?

A prompt tells an AI what to do. A construct tells an AI who to be.

That distinction matters more than it sounds. When you write “review this code for bugs,” you get a generic review. When you give an agent a construct that defines it as a Paranoid Staff Engineer — one who checks for SQL injection, race conditions, LLM trust boundaries, and silent failures, with a specific severity rating system and a defined output format — you get something categorically different.

The first is an instruction. The second is an identity.

The Problem with Prompts

Prompts are ephemeral. You write one, use it, and it disappears into your chat history. There is no version control, no sharing, no building on what others have done. Every team reinvents the same system prompts from scratch.

Constructs solve this by treating agent definitions as artifacts — things that can be published, discovered, forked, improved, and composed. Like open-source software, but for agent intelligence.

What a Construct Contains

A construct typically defines four things:

Identity. Who is this agent? What is its role, expertise, and personality? A construct might define an agent as a “senior product designer with strong opinions about typography” or a “calm incident commander who coordinates response without doing the debugging.”

Principles. How does this agent make decisions? What does it prioritize? A code review construct might say “architecture over formatting” and “never auto-fix architectural concerns.”

Workflow. What does this agent actually do, step by step? A debugging construct might define: reproduce, isolate, hypothesize, test, fix — in that order, never skipping steps.

Boundaries. What does this agent refuse to do? What is out of scope? A QA construct that only reports bugs but never fixes them is fundamentally different from one that reports and fixes.

The Ecosystem

Constructs exist under many names across the ecosystem:

SKILL.md files define capabilities for Claude Code. Garry Tan's gstack system uses twelve of these to create an entire engineering team.

SOUL.md files define personality and values for OpenClaw agents. They encode who an agent is at its core — its beliefs, communication style, and behavioral guardrails.

AGENTS.md files define cross-tool instructions, now standardized under the Linux Foundation.

.cursorrules and .windsurfrules files define behavior for specific IDE agents.

What constructs.sh does is unify all of these under one platform — a place where any construct, in any format, for any harness, can be discovered, forked, and improved by the community.

Why This Matters

AI agents are becoming the primary interface for work. The question is not whether you will use them, but how well they will serve you. The difference between a generic agent and a great one is the construct behind it — the structured thinking of someone who understood the domain deeply enough to teach a machine how to operate in it.

Constructs make that expertise shareable, composable, and buildable. That is why they matter.