Build an AI Agent From Scratch

Build the agent loop that runs the tools the model asks for until it answers.

Build the dispatcher that turns bad tool calls into error results instead of crashes.

Assemble complete tool calls from streamed fragments before anything runs.

Add the stop conditions that end a conversation the model never would.

Run a turn's tool calls in parallel so one hanging tool cannot stall the rest.

Define the record schema and the validator that keeps junk out of the database.

Clean fenced, comma broken model responses into parseable JSON.

Feed validation errors back to the model until its output passes.

Recover every complete field from JSON cut off mid stream.

Build the endpoint that streams tokens to the browser as they are generated.

Stream text to the user while assembling tool calls from the same stream.

Handle multibyte characters split across network chunks so every script renders clean.

Turn an upstream failure into a clean error message instead of a spinner.

Let a dropped client reconnect and continue from the exact word it left off.

Retry rate limited requests with exponential backoff and jitter instead of failing the user.

Bound every request with a timeout so hangs become clean, retryable failures instead of stalls.

Stop hammering a dead provider: open a breaker on a failure burst and probe it back closed.

Serve requests from the secondary provider when the primary fails, without ever billing twice.

Detect responses cut off mid stream and retry them instead of delivering half an answer.