INPUT“Fix the failing test”
promptPrompt + repository context enter the run.
THE AGENT LOOP, MADE VISIBLE
From a prompt
to real work.
nah gives a model a task, a bounded set of tools, and a clear path to act. Every step returns through an explicit loop you can inspect, shape, and stop.
nahAGENT RUN
YOUR TASKprompt + context
MODELdecide next step
TOOLSact in environment
EVENTSobserve the run
control stays in your app01 — 04 / REPEATS UNTIL DONE
THE SHAPE OF A RUN
A loop you can follow.
Each model decision and tool outcome is an explicit part of the run. Nothing disappears behind a one-shot “agent” call.
RUN 8F31 / CODING AGENT IN PROGRESS
MODEL STEPInspect the test and implementation
text → toolThe model requests a read operation.
TOOL RESULTTest output points to a mismatch
tool-resultStructured results become the next model context.
continues with new context
APPROVAL GATEReview a proposed file edit
approvalYour policy decides whether the action can proceed.
FINISHChange verified, summary returned
doneFinal answer and usage are available to your app.
event stream connected5 STEPS · 2 TOOL CALLS · 1 APPROVAL
FOUR MOVING PARTS
Small steps. Clear ownership.
Your application sets the boundaries; the loop carries context from one decision to the next.
01
Give it a task
A prompt, working directory, and system instructions become the run’s starting context. The agent receives only the tools and limits your app gives it.
02
The model chooses
The model can respond with an explanation or request a tool. Each response is one bounded step in a loop your application can observe.
03
Tools do the work
Read, search, edit, and shell tools run in your environment. Mutating actions can pause for your approval before anything changes.
04
Results guide the next step
Tool output returns to the model as context. It can inspect the result, choose another action, or finish with a final answer.
AN EXPLICIT API
Wire the loop into your product.
Use the framework when you want to compose the agent yourself. Stream events into your UI, connect approvals to your policy, and keep the run result in your hands.
Explore the framework agent.tsYOUR APP · TYPESCRIPT
const run = runAgent({model, prompt: "Fix the failing test", system: buildSystemPrompt(cwd), tools: createCodingTools(env, {approveToolCall: askForApproval,}), maxSteps: 12, maxSpendUsd: 0.50,});for await (const event of run.events) {renderRunEvent(event);}const result = await run.result; stream each step to your UIevents → result
ONE RUNTIME, MORE WAYS TO BUILD
Start where you are.
Build a custom agent, use nah as your coding partner, or inspect the work as it happens.
01 / BUILD
Framework
Compose model steps, tools, approvals, and events into your own product.
Build with the SDK 02 / USENAH Agent
Put the same explicit loop to work on real code from your terminal.
Meet the agent 03 / INSPECTStudio
See runs, tool activity, and traces in a workspace built for visibility.
Explore StudioREADY TO BUILD?
Make the loop yours.
Start with the agent loop guide, then shape the run around your product and your users.