How an AI agent fits together
An AI agent isn't one thing. It's a few pieces that work together, and the harness is everything that makes the model into an agent. This is the shape of an agent system and how the pieces connect, the big picture the tool stacks on the homepage fit into. Click any box or node to read the note behind it.
flowchart TB
M["Model (the LLM)"]
subgraph H["Agent harness"]
direction TB
A["Agent"]
Mem["Memory & context"]
T["Tools & environment"]
S["Skills"]
C["Hard & soft controls"]
A -->|uses| T
A -->|uses| S
A -->|reads and writes| Mem
end
E["Evaluation"]
O["Observability"]
A -->|calls| M
E -. tests .-> A
O -. watches .-> A
click C "/wiki/an-agent-harness-is-everything-that-constrains-and-shapes-an-agent"
click Mem "/wiki/agent-memory-persists-context-that-the-model-window-cannot-hold"
click T "/wiki/mcp-is-a-protocol-layer-that-decouples-tools-from-providers"
click E "/wiki/evals-define-what-success-means-for-an-agent"
click O "/wiki/tracing-and-evaluation-serve-different-jobs" Harness
The harness is everything that makes an LLM an agent: the tools and environment it can reach, the hard and soft controls that bound what it can do, and the memory and context it carries. The agent is the model in a loop; the harness is everything around it that turns that loop into an agent.
Model
The agent calls a model (the LLM) to decide what to do next. Which model you pick changes quality, cost, and speed.
Tools & MCP
The agent uses tools to act on the world, reading files, running code, calling APIs. MCP is the standard that lets tools plug in from any provider.
Skills
The agent uses skills, reusable, named procedures it can call on, the way a person reaches for a known technique.
Memory & context
The agent keeps context, what it is working on and what it already learned. Memory holds what the model window cannot, and durable knowledge can live outside the model in a folder of typed markdown the agent reads on demand (OKF).
Evaluation
Before you trust an agent, you test it. Evaluation runs the agent against cases and scores it, so good is a number you track.
Observability
Once it is running, observability records what the agent actually did each run, so you can replay it and find what went wrong.