Two things happened in AI this week, and they contradict each other.

A builder with more than forty agent projects published a post that went viral: Stop building agents. His argument: most of what gets sold as an “agent” is an automation with a language model bolted on, and the automation works better.

The same week, Cognition raised $1 billion at a $26 billion valuation for Devin, an agent that writes software end to end. The market just paid a premium for exactly the thing the builders in the trenches are telling founders to stop buying.

Both can be right. And the gap between them is the most useful question you can ask before you greenlight an agent. It is not “AI or not” (I wrote about that one). It is not “is the output slop” (covered that too). Assume a language model belongs in the process. The question that remains is how much autonomy you hand it. And the default the market rewards, the agent you hand a goal and tell to figure it out, is the most expensive and most fragile version.

Here is the thesis in one line: autonomy is not a feature worth paying more for. In production it is a liability. The market is pricing it as an asset.

What the $26 billion is actually paying for

Cognition’s numbers are not fake. Revenue went from $37 million to $492 million in twelve months. Goldman Sachs, Mercedes-Benz, and the US government are customers. The valuation more than doubled from September. As one analyst who models agent companies for a living put it: the $492M in revenue is real, the multiple isn’t.

What the multiple is buying is autonomy. Devin is not sold as a copilot that suggests lines while you type. It is sold as the agent you hand a task and that hands back working software: it plans, writes, debugs, deploys. The word “autonomous” is what justifies the price.

And it is the exact property that every practitioner says breaks first in a real process.

There is a number that gives the game away. Cognition says 89% of the commits its own engineers make are committed by Devin. That sounds like proof that autonomy works on its own. It is the opposite. Devin performs inside Cognition because Cognition built the scaffolding around it: its own harness, its reviews, its engineers watching every commit. The autonomy you buy in a valuation is not the autonomy you can turn loose in your messy process. What scales is not the agent. It is the agent with the bounds someone put around it.

One decision per step, or a goal and good luck

The line between an automation and an agent is not how much AI is inside. It is how much freedom it has to decide.

A good automation has one decision per step and a clear rule for every branch. You know what it does in each case because you defined each case. An agent gets a goal and the instruction to work it out. In a demo that looks like magic. In a support queue at 2am, with an angry customer and an input nobody anticipated, it is a time bomb.

The builder’s three examples all have the same shape:

• A telehealth founder wanted an autonomous receptionist that “handles everything.” What moved the needle was a workflow that reads intake forms and routes them to the right clinician. Six weeks. Four clinician-hours saved a day. She paid him again the next month.
• A fintech wanted a “fully agentic finance copilot.” What it needed was a script that catches ACH discrepancies before they hit the dispute queue. One model call, the rest plain code. It saved a full ops hire.
• A medspa chain wanted “AI marketing automation.” What it needed was a job that watches the booking system and fires a recovery message when it spots a no-show pattern. Three steps. No agent. Up 14% in revenue last quarter.

None of them is an agent. All of them beat the agent the founder originally asked for.

How do you know which side you are on? The same builder reduces it to a few questions, and they all point at one thing: how much autonomy the process can absorb. Can you draw the flow as clear steps? Then you want an automation. Does it genuinely have more than five branches with unpredictable inputs? Then maybe an agent earns its keep. Is the cost of a wrong answer high? Automation. Will compliance ever look at this? Automation, full stop.

Notice that none of those questions is about the model. They are about the process. Autonomy is not a technical choice you make staring at benchmarks. It is an architecture choice you make staring at your operation.

The autonomous version always sells first

I have been building software since 1990. If thirty-six years taught me one thing, it is that every wave of technology sells its autonomous version first, the magic one, the one that supposedly needs no one. The expert systems of the nineties were going to run the business by themselves. The “anyone can build it without code” tools were going to make programmers redundant. In every cycle, the version that actually stuck and created value was the bounded one, the one with a person deciding the branches that mattered.

Lisa Su, AMD’s CEO, said it plainly this week in a commencement speech at MIT, and it is worth reading slowly: for everything AI can do, it cannot decide which problems are worth solving, it cannot make the hard judgments when the data is not there, and it cannot take responsibility for the outcomes. Her close lands directly on this: we do not need more people who know how to use the tool, we need people who know how to apply it to solve big problems.

Deciding how much autonomy to give an agent, where you let it run and where you stop it, is exactly the judgment Su says stays human. It is not a temporary limit that disappears when the model gets better. It is the responsibility you do not delegate.

What IQ Source does about this

When a company asks us for an agent, the first question is not which model to use. It is where, exactly, in this process, the model should decide on its own, and where it should only execute steps we defined.

AI Maestro is the discovery where that line gets drawn, process by process. Two months mapping the real operation, not the org chart, to assign each step its level of autonomy: this one the model decides, this one a rule decides, this one a person decides. It ends in an AI Opportunity Score and a Go/No-Go gate that, for many processes, recommends the bounded automation instead of the agent the client came in asking for.

It is the same discipline as the day an agent invents a step between the ones you control and bills a customer zero dollars: the problem is almost never the model, it is how much rope you gave it without looking. Bounding is not distrust of AI. It is putting AI where it performs.

Cognition is worth $26 billion because it built the scaffolding that makes Devin’s autonomy useful. Your company does not need to buy that autonomy. It needs to decide, step by step, how much its operation can absorb. That decision is the asset. The loose agent is the liability.

Next time someone on your team proposes “let’s build an agent,” ask one thing before you approve the budget: at which exact step do we need the model to decide on its own, and why isn’t a rule enough there? If nobody has the answer, you do not need an agent yet. You need to draw the line.

Draw your agent’s autonomy line