diff --git a/topics/ai-agent/index.md b/topics/ai-agent/index.md
new file mode 100644
index 00000000000..f0d5deb6aae
--- /dev/null
+++ b/topics/ai-agent/index.md
@@ -0,0 +1,17 @@
+---
+aliases: agent-harnesses, ai-agent-harness, llm-harness
+display_name: Agent Harness
+short_description: Runtime infrastructure that connects LLMs to real-world tools, memory, and execution environments.
+topic: agent-harness
+related:
+  - ai-agent
+  - orchestration
+  - tool-use
+  - mcp
+  - llm
+---
+<!-- markdownlint-disable MD041 -->
+
+An **agent harness** is the runtime layer between a large language model and the outside world. It manages the agentic loop — prompt the model, parse tool calls, execute tools, feed results back — while enforcing permissions, compressing context, and handling error recovery. The term gained traction after Anthropic's Claude Code source was studied, revealing the internal architecture that turns a text-generating LLM into a capable software-engineering agent.
+
+Agent harness projects include full runtime environments (LobeHub, oh-my-agent, HexAgent, Utah), agent harness frameworks for Python and TypeScript (water, Chorus), code-quality agents (desloppify), portable skill systems (cc-harness-skills, oh-my-agent), research surveys (LLM-Agent-Harness-Survey), and autonomous systems with persistent memory and orchestration (Adam, Sibyl). The pattern is characterized by a tool registry, a permission system, context management, and often multi-agent coordination.