Four Ways to Build an Agent
Agent frameworks, harnesses, runtimes and builders solve different problems, and choosing the wrong kind is the costliest mistake in an agent project. This map explains the four, helps you pick, and opens a series that takes the main systems apart one by one.
1. Why this series
- Everyone says "agent". They mean four different things.
- A framework gives you parts to assemble; a harness is a working agent you configure; a runtime keeps agents running in production; a builder lets non-developers design one visually.
- Picking the wrong kind is the most common, most expensive mistake in agent projects: writing orchestration code for a task an existing harness already does, or putting a laptop harness in front of thousands of users.
- What each article does.
- Takes one real system apart: case file, architecture diagram, one task traced end to end, building blocks, a configuration example, when to choose it, and how it compares.
- Harness articles all trace the same task (fix a failing test and open a pull request) so you can compare them directly.
2. The four kinds
🗺️ The four kinds of agent system (tap one)
Framing adapted from an infographic by Denis Panjuta (third party); descriptions and examples checked against each project where we could.
3. Which kind do you need?
🧭 Which kind do you need?
A rule of thumb, not a law. Real systems often combine kinds: a framework agent deployed on a runtime, or a harness built with a framework.
4. How they fit together
🥞 How the kinds stack
Our simplification. The Agent Layer series covers the infrastructure underneath all four.
- Three examples.
- A support bot drawn as a flow in a framework, deployed on a runtime for thousands of customers.
- A developer using a harness (Claude Code, Codex) on a laptop: no runtime needed.
- A sales team wiring a lead-qualification agent in a builder (n8n, Zapier); the builder runs it.
5. The eight layers inside every harness
🏛️ The reference stack used in every teardown
Each harness article draws its real components into these eight bands and traces one task through them, so you can compare systems layer by layer.
6. What is published, and what is next
- Wave 1 (now): this map, the hub page and five harnesses.
- Claude Code · Codex · Gemini CLI, Antigravity and Jules · Deep Agents · Hermes Agent.
- Wave 2: frameworks.
- OpenAI Agents SDK · Google ADK · LangChain/LangGraph · CrewAI · Claude Agent SDK and Microsoft Agent Framework.
- Wave 3: runtimes.
- Amazon Bedrock AgentCore · LangSmith Deployment · Temporal · Vertex AI Agent Engine and Azure AI Foundry Agent Service.
- Wave 4: builders.
- n8n · Zapier Agents · Microsoft Copilot Studio · Make.
7. The series
Agentic AI System Designs: four kinds of agent system, one article each
See them all on one page: Agentic Systems at a Glance.
8. Sources
- Sources
- Infographic "Agent framework vs. agent harness vs. agent runtime vs. agent builder" by Denis Panjuta (third party; framing only)
- Inside the Agent Layer: the infrastructure under all four
- Agentic Systems at a Glance (hub)