From chatbot to agent: the five levels of autonomy of agentic systems, per Google, and which one Charla already operates at today.
Yohan Consani · Published
Artificial intelligence is changing. For years the focus was on models that respond to a prompt: answering a question, translating text, generating an image. Now we're living through a shift: from AI that merely predicts or creates content to software that can solve problems and execute tasks autonomously. This is the era of AI agents.
In its whitepaper Introduction to AI Agents, Google defines an agent as, in essence, “a language model in a loop, with tools, to accomplish a goal.” But not all agents are equal: the document organizes these systems into five levels of capability, each built on the last. Most tools on the market still stop at Level 1.
Agents are the natural evolution of Language Models, made useful in software.Google · Introduction to AI Agents
The five levels of agentic systems
Level 0 : The core reasoning system
The starting point: the language model alone, with no tools, memory, or access to the world. It can explain concepts and plan in depth, but it's “blind” to any fact outside its training data: it doesn't know last night's score or today's inventory.
Level 1 : The connected problem-solver
Here the model gains “hands”: it connects to external tools and data. With RAG (retrieval-augmented generation) it looks up current information before answering and grounds the response in facts. This is the level of assistants that answer one-off questions from a knowledge base, where most enterprise copilots operate.
Level 2 : The strategic problem-solver
The leap is “context engineering”: the agent starts planning multi-step tasks, actively selecting and packaging the right information for each step of the plan. It moves from running isolated tasks to driving processes, even proactively.
Level 3 : The collaborative multi-agent system
Instead of a single “super-agent,” a team of specialists works in concert, like a human organization. One agent coordinates, others execute, and patterns like “generator and critic” (one agent creates, another evaluates) raise quality. Agents begin to use other agents as tools.
Level 4 : The self-evolving system
The highest level: the system identifies gaps in its own capabilities and dynamically creates new tools (or even new agents) to fill them. It stops using a fixed set of resources and starts expanding it on its own, becoming an organization that learns and evolves.
What level is Charla at today?
Our honest assessment: Charla operates today at Level 3, with a foot in Level 4. It goes well beyond a chatbot querying a knowledge base (Level 1) or a single agent that plans (Level 2). In practice, Charla already delivers:
- Knowledge that becomes insight: it understands a vast base (documents, spreadsheets, and images), serves customers, integrates with systems and APIs, and automates entire workflows.
- From unstructured to structured: it extracts data from messy sources into the format you want (spreadsheet, presentation, or database) and compares and validates documents against guidelines.
- Natural-language data analytics (C3QL): it queries large databases and returns analytical reports for technical teams.
- Machine-learning models as tools: forecasting, recommendation, and optimization in service of your KPIs.
- Charlie: an agent that creates other agents, interactively with the user.
- Steve: continuous self-evaluation, judging answers against a gold standard.
Why Level 3? Because Charla isn't a lone model with tools: it's a team of specialized agents working together. The answering Charla operates alongside Steve, which evaluates and critiques every answer (the classic “generator and critic” pattern) and alongside agents that compile knowledge and prepare the data context. That division of labor is the hallmark of Level 3.
And why a foot in Level 4? Because Charlie already does what Google describes as the heart of the self-evolving level: creating new agents. Today this happens in a guided way, with a person in command (human-in-the-loop): a safe, auditable path toward full autonomy.
The goal: Level 4
Our goal is clear: to take Charla to full Level 4, with multi-agent topologies that coordinate themselves, long-term memory across conversations, and systems that build their own tools to fill gaps. A fleet of agents that learns, evaluates itself, and expands continuously. We're building an agentic organization.