Levels of Agentic Behavior

A note on words. Throughout this piece I talk about workflows. A workflow is the steps necessary to accomplish a task. I call each step a node, and the lines connecting the nodes are edges.

When thinking about agentic AI systems, I think in levels of agentic behavior. The levels are determined by how much we trust the system. Rather than grading particular AI abilities or technical techniques that might come and go, trust gives us a durable model.

This mirrors how we handle other autonomous things in our lives. A five-year-old might be physically capable of operating a stove, but we don't trust them to cook dinner unsupervised. A teenager might understand the mechanics of driving, but that doesn't mean they're ready for a cross-country road trip.

The self-driving industry gives us the parallel. The Society of Automotive Engineers doesn't classify autonomous vehicles by their technical implementation, whether they use cameras, LiDAR, or some proprietary sensor fusion. It classifies them by how much human intervention is required. The scale is about trust, not architecture.

So: agentic AI along the same spectrum. As we move from one level to the next, we hand off a larger percentage of control.

Level 0: Rule-Based Workflows, 100% Human Control

Level 0 is what Zapier popularized, an "if this, then that" workflow builder. Everything about the workflow is deterministic. A node might call an API and retrieve dynamic data, but what happens with that data is pre-determined by the workflow you set up.

SAE Level 0: No Driving Automation. The driver has complete control of the vehicle, and no automation features are active.

When tools like Zapier first emerged, they felt revolutionary. We could automate tasks without writing code. Looking back, it's closer to what an automatic transmission was to a manual stick shift. An easier user experience, still deterministically controlled by the human.

At this level, humans define every step, every condition, and every action. The system executes them faithfully. There's no ambiguity, no interpretation, and no autonomy.

Why you'd stay here. Repeatability beats flexibility. Nightly data jobs, compliance workflows, anything where predictability is the point and every run looks like the last one.

Level 1: The LLM Controls the Prompt Nodes

At Level 1 we start introducing LLM calls into the workflow, though the overall structure stays deterministic. Parts of your workflow start to feel like they can think. A thinking node can tell us whether a user sounds upset. If they are, we go down one branch. If not, another.

SAE Level 1: Driver Assistance. The vehicle can assist with either steering or acceleration and braking, but not both at once.

This is the cruise control in your first car. You give it basic flexibility to speed up and slow down inside parameters you set. At first we didn't even trust these systems to slow down on their own.

Level 1 is our first tentative step toward giving up control. The system gains limited flexibility in how it processes information, and everything else stays controlled and structured. This was possible with traditional AI models before LLMs, but it was much harder to build a model to classify a photo as hotdog or not hotdog than it is to use an LLM today.

Level 1 agents became practical around the GPT-3 era. What changed wasn't the capability, it was the activation energy. Rather than hiring a team of PhDs, collecting data, and working for a year to build a custom model, you could plug into an LLM. That dropped the barrier to putting a probabilistic step inside an otherwise deterministic workflow.

Level 2: The LLM Controls the Tools It Can Use

At Level 2 the LLM decides where to pull data from and which tools to use to achieve predefined tasks. The agent chooses among approved APIs. It cannot create new goals or establish its own feedback loops.

SAE Level 2: Partial Driving Automation. The vehicle can handle both steering and acceleration and braking under certain conditions, but the driver must stay engaged and ready to take control.

If you squint, you might mistake this for a full agent. It's the same way lane keeping made a gentle highway curve feel like self-driving. Then the turn gets sharp, the system fails, and you take over.

Tool use makes the intelligent node powerful inside a tightly controlled workflow. The AI decides which tools to use, inside a workflow a human carefully defined. The moment it meets a scenario outside its programming, a human has to step in.

Guardrails you need here. Tool contracts with a fixed schema, rate limiters, and a safe execution sandbox.

At Brevy we mostly operate at Level 2. Everything past this point in the article is where the future takes us. The higher levels are in use at other companies and don't make sense for us yet.

Level 3: The LLM Controls the Workflow

At Level 3 you move away from rigid, pre-defined workflows. You provide the possible nodes, the available tools, and the desired outcome. The AI determines the path. It draws its own edges.

SAE Level 3: Conditional Driving Automation. The vehicle can handle most driving tasks under specific conditions, but the driver must be ready to take over if the system fails or the situation gets too complex.

When a system starts determining its own workflows, it feels like AGI is here and all of your work can be handed over. That feeling is riding in a Tesla with Full Self-Driving. It works beautifully 95% of the time, and if you weren't there for the other 5% the outcome could be disastrous.

This is roughly where agentic AI is today. The challenge is keeping the agent marching toward progress as the workflow gets long. Build one of these and you will meet the infinite loop, where the agent loses sight of the goal and devolves into chaos. Much of the current research is about keeping the loop running longer while making sure it's still making progress.

Today's duct tape. Budgeted token windows, evaluators, and synthetic tests.

Level 4: The LLM Controls Beyond the Workflow

At Level 4 agents become substantially more useful. They work across long time horizons and across tools beyond their original domain. You no longer trigger every run or specify every tool call.

SAE Level 4: High Driving Automation. The vehicle can perform all driving tasks under certain conditions without driver input, though the driver may need to take over in specific situations.

Level 4 is where Waymo sits. This is where you take the human out of the driver's seat. The AI does everything, inside a carefully defined domain. For a car that might be a specific region of San Francisco or Austin. For agents it might be a small set of tools and business functions you have thoroughly tested.

An agent you tuned to write code isn't suddenly allowed to send sales emails. The boundaries are still there. Inside them, the system runs with real autonomy. It can schedule tasks, monitor systems, and decide without constant oversight.

Why it changes the game. Humans leave the hot loop and move to exception handling. Picture an agent that scans your calendar every fifteen minutes and books the ride to your next meeting without being asked.

The narrow domain is the point, not a limitation. Bounding the domain is what buys you the reliability that makes autonomy safe.

Level 5: The LLM Controls Across Work

At Level 5 little has changed architecturally from Level 4. The systems have simply become reliable enough that we trust them across much larger surface areas. They reach your work tools and pursue tasks over long horizons and wide domains.

SAE Level 5: Full Driving Automation. The vehicle can handle all driving tasks under all conditions, and no human intervention is needed.

Imagine Waymo driving anywhere, with no restriction, problem-solving its way through whatever it meets. A Level 5 agent needs the same: a working understanding of where to find resources, where to put information, and how to treat a human as a collaborator.

Consider a product management agent that pulls user analytics, writes specs, designs mockups, sends them to test users, builds an MVP, shares updates with the team, and iterates on the feedback, with minimal human involvement.

At this level the agent is a collaborator rather than a tool. The boundaries between systems become transparent to it, and it moves between them in service of a higher-level goal.

What the human does. Set objectives, watch the numbers, audit occasionally. That's it.

How to Know You're Ready to Level Up

Four questions before you move up a level.

1. Error budget.Can you afford a misfire? What's the worst case if the agent gets it wrong?
2. Observability.Do you log every tool call and every decision? Can you trace exactly what happened?
3. Rollback.If the agent goes rogue, can you kill it and patch quickly?
4. Regulation.The narrower the domain, the easier the sign-off.

Master one level before reaching for the next. A Level 2 system that reliably picks the right tool inside a real domain is worth far more than a Level 4 system that demos well and fails unpredictably in production.

The whole scale is about trust, and trust is earned the same way it is with a car that drives itself. Reliability you can demonstrate, and constraints you chose on purpose.