Nobody asks about Agent Consciousness Anymore
Last year, talking about AI consciousness was fashionable. ChatGPT sounded human, called itself “I,” and could write remarkably convincing passages about feelings, doubt, and its own existence. That was enough to trigger endless speculation about whether language models might already be conscious.
I considered most of that debate nonsense from the beginning. Linguistic brilliance is not evidence of subjective experience. Today, a certain degree of sanity seems to have returned.
The reason is simple: we are building and using agents. We don't chat with them.
Agents demystify language models
In a chat, we mainly see the answer. With an agent, we see the work.
We see which files it opens, which tools it calls, which assumptions it makes, and where it takes a wrong turn. We watch it forget requirements, get stuck in loops, overlook errors, or pursue a misunderstood objective with impressive determination.
That clears your head. A chatbot can create the illusion of another mind. An agent exposes the machine.
And that machine is anything but trivial. It can analyse code, research information, edit files, run tests, and handle complex tasks across many steps. But its performance does not emerge from a mysterious digital inner life. It emerges from an architecture:
A model, context, tools, memory, loops, permissions, and controls.
The interesting questions are technical now
The moment a model is expected to act, sounding convincing is no longer enough.
What matters is:
- Did the agent actually understand the task?
- Does it have the context it needs?
- Which tools is it allowed to use?
- Which files and systems may it modify?
- Can it verify intermediate results?
- Does it notice when it is stuck?
- When should it stop or ask a human?
- How do we prevent a small error from propagating through twenty autonomous steps?
These questions are less dramatic than “Does the machine have a soul?” But they determine whether an agent is useful, expensive, dangerous, or simply annoying.
A conscious system that works unreliably would be useless for almost every practical task. A non-conscious system that performs complex work reliably can transform entire professions.
An agent is not merely a chatbot with tools
A good agentic system needs more than a tool button attached to a language model.
It needs a clear objective and a stable working context. It must distinguish binding information from assumptions. It needs suitable tools, limited permissions, and visible state. It must verify results, detect failures, and recognise when further action no longer makes sense.
Most importantly, it needs an architecture that does not merely generate errors, but catches them.
That is the real core of agentic engineering. The goal is not to hope that the model somehow becomes magically smarter. The whole system must be designed so that the model’s strengths become useful and its weaknesses remain manageable.
The consciousness question was the wrong shortcut
The fascination with supposedly conscious LLMs partly arose because the systems initially felt like magic. They could suddenly do things we associated with thought. The tempting conclusion was: if they think like us, perhaps they also experience like us.
That conclusion was never convincing.
A language model can talk about pain without anything hurting. It can describe itself without possessing a stable self-model. It can formulate empathy without feeling anything.
I explain in more detail why I do not consider current LLMs conscious — and why I still think artificial consciousness may be technically possible — on raphaelkirchner.com: Agents Ended the Nonsense About Conscious LLMs.
For agentic work, the issue is simpler anyway. We do not need to know whether some tiny spark of experience exists somewhere inside the system. We need to know whether it completes its task correctly.
We are finally discussing the right things
The consciousness question was not solved. It was displaced by more important questions. That is progress.
We now discuss which tasks agents should receive, which context they need, which doors we open for them, and how we control what they do once they enter. Not what the system claims to be. What it actually does.