Building a coding agent with the OpenAI SDK and .NET

Coding agents aren’t one giant prompt. They’re a loop with tools: think, read files, edit slices, search the repo, and repeat until the task is done.

The minimal pipeline

A useful .NET agent usually exposes four capabilities:

• —Plan — break the request into verifiable steps.
• —Read — load only relevant files (avoid context dumps).
• —Edit — small, reviewable patches—not full rewrites.
• —Grep — semantic or text search to find symbols and usages.

public interface ICodingAgentStep
{
    Task<AgentObservation> RunAsync(AgentState state, CancellationToken ct);
}

Orchestration in C#

With the OpenAI SDK (or another provider), keep state outside the model: step history, pending diffs, and stop criteria. The LLM suggests the next action; your code validates and executes.

When the agent fails, prefer fixing the loop (better observation, scope limits) over growing the prompt.