Anatomy of an
AI Agent System
See how the model, information, tools and operating software work together during a run. Select any part to open its detailed explainer.
Select any part
What each part contributes
The model thinks and acts, while the surrounding system gives it direction, information and a way to work.
At the centre
What guides and informs the model during a run
Instructions
The rules you write in plain English covering the goal, the method, the limits and examples of good work.
Compare no-code and code →Context
The working information available to the model at that moment, such as your request, relevant files, conversation history and tool results.
Compare no-code and code →Tools
The actions it can take in the real world, like searching a folder, updating a record or sending an email.
Compare no-code and code →Memory
Information saved outside the model and brought back into context when relevant, such as preferences, previous decisions or unfinished work.
Compare no-code and code →Feedback
Observations returned during the run, such as tool results, errors or human input, which help the model choose its next step.
Compare no-code and code →The operating layer and quality check
Harness
The software that assembles the context, calls the model, executes tools, manages state and permissions, repeats the cycle, and stops when the task is complete or a limit is reached.
Compare no-code and code →Evaluation
A systematic check of completed runs using expected answers, scoring rules or human review to see how reliably the agent performs.
Compare no-code and code →