Claude Is Becoming More Than a Chatbot

For most of the generative AI era, the dominant interaction has been remarkably simple.

You ask a question.

The model gives you an answer.

The answer might be a paragraph, an analysis, some code, an image or a recommendation. The quality has improved enormously, but the basic relationship has remained recognisable: human request, machine response.

Anthropic’s development of Claude increasingly points towards something different.

The important change isn’t simply that Claude is becoming better at generating outputs. It is becoming more capable of reasoning through problems, using tools, taking actions and continuing work across longer sequences without requiring a human instruction at every step.

That changes the unit of value.

Instead of asking how good an individual answer is, we can increasingly ask how much useful work the system can complete before it needs us again.

The frontier is no longer only about the quality of individual outputs. It is increasingly about the length and complexity of the work a model can successfully perform.

Reasoning changed what we could ask#

Earlier this year, I wrote about how OpenAI o1 changed what businesses should ask of AI.

The significance of reasoning models was not merely that benchmark scores improved. They demonstrated that giving a model more computational effort on a problem could improve its ability to handle work that required several stages of thought.

Anthropic pushed a similar idea in February with Claude 3.7 Sonnet, which it described as its first hybrid reasoning model. Users could choose between relatively immediate responses and extended thinking, while API users could control how much thinking the model performed.

That distinction matters because not every problem deserves the same amount of intelligence.

Writing a short summary might need very little deliberation. Debugging a difficult software problem might need much more. A model that can vary the amount of reasoning it applies begins to look less like a fixed software function and more like a flexible source of cognitive effort.

But reasoning by itself still leaves the model largely inside the world of answers.

The bigger shift happens when reasoning is connected to tools.

Claude Code made the direction clearer#

Anthropic launched Claude Code alongside Claude 3.7 Sonnet in February as a research preview.

Its description was revealing: developers could delegate substantial engineering tasks to Claude directly from their terminal.

The word delegate is important.

Traditional software usually waits for explicit instructions. Even sophisticated automation tends to follow a predefined sequence: when this happens, do that; call this system; move this data; execute this rule.

An agentic system can operate differently.

You can give it a goal, provide access to relevant tools and allow the model to work out some of the intermediate steps itself.

Automation tells software how to do the work. Delegation tells AI what you want accomplished.

That doesn’t mean the system has unlimited autonomy. Claude Code still operates within permissions, tools, context and constraints. But the human can increasingly specify the desired outcome rather than every action required to reach it.

That is a fundamentally different relationship with software.

Reasoning and action are starting to combine#

In April, I described AI agents as a movement from answers towards actions.

The underlying loop is straightforward:

Think → Act → Observe → Think again.

A model reasons about what to do, uses a tool, observes what happened and then decides what to do next.

Anthropic’s May release of Claude Opus 4 and Claude Sonnet 4 made this increasingly explicit. The models could use tools during extended thinking, allowing Claude to alternate between reasoning and tool use. They could also use tools in parallel, while Anthropic emphasised improvements for complex, long-running tasks and agent workflows.

This is important because useful work rarely consists of a single perfect action.

A developer trying to fix a software bug might need to inspect several files, understand the architecture, search for a relevant dependency, modify code, run tests, interpret the failures, make another change and test again.

A research task might require finding sources, comparing evidence, identifying contradictions, calculating something and producing a final document.

A business process might involve retrieving customer information, checking policy, updating a system, sending a communication and escalating an exception.

The value comes from the sequence.

If a human has to instruct the model after every step, the AI is useful but still heavily supervised.

As models become capable of completing more steps correctly before requiring intervention, the economics begin to change.

Sonnet 4.5 extends the length of the work#

By September, Anthropic’s direction was even clearer.

Claude Sonnet 4.5 arrived on 29 September, with Anthropic positioning it particularly strongly around coding, computer use and complex agents.

At the same time, Claude Code gained checkpoints that allow developers to rewind changes, while Anthropic released the Claude Agent SDK so developers could build agentic systems using the same underlying infrastructure that powers Claude Code.

Anthropic also described work on context management, permissions and coordinating subagents across longer-running tasks.

These details can sound like developer tooling, but they point to a much broader problem frontier AI companies are trying to solve.

How do you allow an intelligent system to work for longer without losing control of what it is doing?

That problem becomes more important as the amount of delegated work increases.

If a model produces one paragraph, the cost of failure is usually limited.

If it performs fifty connected actions across several systems, the consequences can be very different.

Longer work therefore requires more than better intelligence. It requires mechanisms for permissions, observation, recovery, context and human intervention.

Autonomy and control have to develop together.

This isn’t only an Anthropic direction#

Anthropic is a useful lens because Claude Code makes the progression unusually visible, but the wider frontier is moving in the same direction.

OpenAI described agents in March as systems that independently accomplish tasks on behalf of users when it introduced its Responses API and agent-building tools.

Then in July, ChatGPT agent combined reasoning with a visual browser, text browser, terminal and other tools. The system could move between reasoning and action while carrying out multi-step tasks.

Different companies will make different product and architectural choices, and leadership on individual benchmarks will continue to change.

Businesses should therefore be cautious about building their AI strategy around whichever frontier model happens to lead a particular evaluation this month. As I argued when discussing why you can’t pursue every AI opportunity, prioritisation matters more than chasing every new capability.

The more important question is what these systems are collectively becoming capable of doing.

And the direction looks increasingly clear.

Generation is being combined with reasoning.

Reasoning is being combined with tools.

Tools allow actions.

Sequences of actions allow larger units of work to be delegated.

Delegation changes the economics of AI#

This is where the technological development becomes commercially interesting.

Imagine a capable employee using an AI assistant to draft an email.

The human decides what needs to happen, gathers the relevant information, asks the AI for a draft, reviews it and sends it.

Now imagine the AI can retrieve the relevant customer history itself, identify the issue, check the company’s policy, draft the response, update the CRM and prepare the next action for approval.

The model hasn’t merely become better at writing.

The boundary of delegated work has expanded.

That matters because human attention often sits between tasks.

We gather information before doing something with it. We move information between systems. We interpret one output before initiating the next step. We monitor processes. We decide when something needs escalating.

If AI can safely perform longer sequences of those activities, it begins to affect workflows rather than individual tasks.

That connects directly to my argument that AI won’t transform work unless the workflow changes.

As agents become capable of participating in larger units of work, organisations can redesign processes around a different assumption: the human may no longer need to sit between every step.

The question becomes where human involvement creates value rather than where software historically required it.

More autonomy makes context more important#

There is an important constraint on all of this.

The more work we delegate, the more the AI needs to understand about the environment in which it is operating.

A chatbot answering a generic question can rely largely on what it learned during training.

An agent acting inside a company cannot.

It needs to know which customer is involved.

Which policies apply.

What happened previously.

Which systems contain authoritative information.

What it is allowed to change.

What requires approval.

What constitutes an exception.

Who should be notified.

What the organisation is actually trying to achieve.

This is why organisation-specific context becomes so important.

The greater the autonomy, the more consequential the context.

A model with incomplete context might produce a slightly worse answer.

An agent with incomplete context might take the wrong action.

As capability expands, connecting AI to the right organisational knowledge becomes more valuable, not less.

Delegation also changes the role of the human#

There is a tendency to imagine autonomy as a binary choice.

Either the human does the work or the AI does it.

In practice, the interesting territory is likely to sit between those extremes.

A human might define the objective while an agent works out the sequence.

The agent might handle routine cases while a person manages exceptions.

The AI might prepare and execute a series of reversible actions but require approval before a consequential one.

A person might review outcomes rather than supervise every intermediate step.

Different work will justify different levels of autonomy.

The appropriate boundary will depend on reliability, consequence, reversibility, regulation, cost and the quality of the available context.

This makes agent design partly an organisational design problem.

Businesses will need to decide not merely what AI can do, but what it should be allowed to do and where human judgement still creates distinctive value.

Benchmarks tell us only part of the story#

Frontier-model competition naturally produces attention-grabbing benchmark results.

Those results matter. They tell us something about the underlying capability of the models.

But as AI becomes more agentic, individual benchmark scores capture a smaller part of what businesses care about.

A model might be slightly better at coding but worse at following instructions over a long task.

Another might reason brilliantly but struggle to use a particular tool reliably.

A system might perform well in a controlled evaluation but become expensive when operating across hundreds of steps.

A technically weaker model might be more useful inside a particular workflow because it is faster, cheaper, easier to control or better connected to the required systems.

The commercial question is therefore not simply:

Which model is smartest?

It is:

Which system can reliably complete the work we need, with the economics and controls we require?

That is a much more useful question for businesses.

The unit of AI capability is getting larger#

The first wave of generative AI made individual outputs dramatically easier to create.

A paragraph.

An image.

A summary.

A piece of code.

A recommendation.

Reasoning models expanded the complexity of problems those outputs could address.

Tool use allowed models to interact with information and systems beyond the prompt.

Agents connect those capabilities across sequences of work.

The natural consequence is that the unit of AI capability becomes larger.

Instead of generating one thing, the system can potentially complete a task.

Then a process.

Eventually, perhaps, substantial parts of a role or business function.

We should be careful about assuming how quickly that progression will happen. Today’s agents still fail, lose context, misunderstand instructions and require supervision. Long-running autonomy remains a difficult technical problem.

But the direction matters even before the technology is perfect.

Because every increase in the amount of useful work that can happen between human interventions changes what businesses can sensibly consider delegating.

That is why Claude becoming more than a chatbot matters.

Not because every company should suddenly deploy Claude Code.

Not because Anthropic will necessarily lead every model benchmark.

And not because fully autonomous digital workers are about to replace organisations.

It matters because the interface between people and intelligence is changing.

We are moving from telling software exactly what to do, towards increasingly being able to describe what we want accomplished and allowing intelligent systems to work through more of the path themselves.

That creates a different question for every organisation thinking seriously about AI.

Not simply:

What can this model answer?

But:

How much more useful work can we safely give it and walk away from?