confused-ai

confused-ai

confused-ai is a TypeScript agent framework built around one stable install story: start with a single package, ship one useful agent, then layer tools, retrieval, sessions, serving, orchestration, and production controls without changing frameworks midway through the project.

One quick example

import { agent, tool } from 'confused-ai';
import { z } from 'zod/v3';

const getQuote = tool({
	name: 'get_quote',
	description: 'Return a stock quote for a ticker symbol.',
	parameters: z.object({ symbol: z.string() }),
	execute: async ({ symbol }) => ({ symbol, price: 927.5, changePct: 1.4 }),
});

const financeAgent = agent({
	name: 'finance-agent',
	model: 'gpt-4o-mini',
	instructions: 'Use the tool to answer market questions in one concise sentence.',
	tools: [getQuote],
});

const result = await financeAgent.run("What's NVDA trading at today?");
console.log(result.text);

The intended feel is simple: plain TypeScript, one explicit capability at a time, and a direct path from small prototype to production-ready runtime.

What it is for

Use confused-ai when you want to build one of these shapes from the same public API surface:

• a single agent that answers, summarizes, or classifies
• a tool-backed assistant that reads live application data or triggers side effects
• a retrieval-backed system that answers from documents or indexed knowledge
• a served application with sessions, resilience, and observability
• a multi-agent workflow with delegation, routing, or explicit reasoning steps

The design goal is not to force every feature on day one. The design goal is to let the first useful version stay small while keeping a direct path to a larger system.

Three primitives

Primitive	Use it when
Agent	one model-backed worker can handle the task
Team	specialists should coordinate or delegate work
Workflow	the execution path should be staged, deterministic, or branching

These three shapes cover most systems in the framework. The difference is not branding. The difference is how control flows through the application.

How to approach the framework

The cleanest adoption path is:

• Start with one agent and one successful run.
• Add one missing capability at a time, usually a tool, a session store, or retrieval.
• Add runtime surfaces such as HTTP serving, scheduling, evaluation, or resilience only after the base behavior is correct.

That order matters because it keeps the model behavior understandable before infrastructure complexity gets involved.

Public package story

The public install story is intentionally simple.

Import path	Use it for
confused-ai	core agent authoring, composition, and common entry points
confused-ai/session	session stores and continuity
confused-ai/serve	HTTP runtime
confused-ai/tool	MCP and broader tool infrastructure
confused-ai/orchestration	teams, supervisors, roles, and tasks
confused-ai/reasoning	explicit reasoning steps and events
confused-ai/scheduler	scheduled jobs and run history
confused-ai/observe	traces, metrics, and evaluation workflows
confused-ai/adapters	infrastructure adapters and bindings
confused-ai/guard	runtime control primitives such as circuit breakers

Avoid internal @confused-ai/* package imports in application code and public documentation. Those paths describe the monorepo layout, not the intended consumer API.

Core building blocks

The framework stays understandable if you think in layers:

• Agents are the unit that owns instructions, model selection, tools, and runtime behavior.
• Tools are the bridge to live data, side effects, and application-specific capabilities.
• Sessions, memory, knowledge, and storage add continuity or external context.
• Serving, scheduling, and orchestration control how and when the agent runs.
• Observability, budgets, approvals, and resilience turn a useful agent into an operable system.

Each layer is optional. Most real projects only need a subset.

Capabilities

Capability	What it gives you
Tools	explicit boundaries for live data and side effects
Sessions	continuity across turns
Memory	retained facts and selective recall
Knowledge	retrieval-backed answers from indexed content
Storage	durable state around the agent
Serve	HTTP runtime for real applications
Orchestration	teams, supervisors, roles, and routing
Reasoning	explicit reasoning loops when the task needs them
Scheduler	time-based execution for reports, digests, and automation
Observe	traces, metrics, and evaluation workflows
Guardrails and HITL	validation, approvals, and policy-driven runtime control

Recommended reading order

If you are new to the repo, follow this order:

• docs/guide/introduction.md for the mental model and product story.
• docs/guide/getting-started.md for the first implementation path.
• docs/examples/index.md for runnable examples by difficulty.
• docs/guide/ pages for capability-specific guidance.
• docs/api/ pages for a compact public API map.

What to build first

The first milestone should be boring on purpose:

• one prompt
• one model
• one agent
• one verified output

Once that path is correct, the rest of the framework becomes a set of focused additions rather than a wall of concepts to learn up front.