What Are Agents?

Agents are AI systems that can do work independently in the background, performing long-running multi-step tasks, without requiring continuous user input each step.

This is contrast to chatbots like ChatGPT, that wait for your next prompt after each step or response.

The goal with agents is to create a "drop-in remote worker" that you can interact with like you do an employee.

Agents Vs Chatbots: Paradigm Shift Or Evolution?

I've attempted to distinguish agents from chatbots in two ways:

1. They perform work in the background.
2. They perform multiple steps of work to complete a task without additional input.

However, the line is blurring as our favorite chatbots get upgrades. We're now seeing chat systems:

1. Run more "reasoning" (internal monologue) before responding -- do more work in the background before responding.
2. Perform multi-step workflows quickly, like accessing files, then running some code, then summarising with a response.

In this sense the name "agent" is a rebranding of chatbot. It is however, a name that makes sense; As these AI systems unlock the ability to do more work on their own, they gain more agency.

Examples of Agents:

• OpenAI's Deep Research -- Performs dozens of web searches, synthesizes them, and then produces a final response.
• OpenAI's Codex -- Writes code in its own cloud environment and comes back to you later with changes to merge into your codebase.

These examples demonstrate that agents are already capable of generating significant value for certain generic tasks.

Current Limitations Of Agents

General purpose agents still make mistakes:

• They use the wrong tools in the wrong places.
• They misinterpret your instructions, especially over long and complex tasks.
• They can get stuck in loops.

These are all obvious problems, and solving them unlocks a lot of value that companies are willing to pay for. As such, they're getting a lot of attention. So its likely we will see them addressed over time, and you can count on us to keep you updated.

"Everything AI Agents" Vs Focused AI Agents

Current situation: General purpose agents can already accelerate lots of work, but they face issues outlined above.

As nice as it might be to have one agent to rule them all, we don't have them yet.

Related to this, focus is every startup's superpower.

By focusing on a specific set of sub-tasks, teams have already managed to create problem/domain-specific agents that outperform the generic ChatGPTs of the world.

We often hire people not because they're generally smart, but because they have expert knowledge, they understand best practices and workflows, and they're familiar with the tools.

This type of expertise is not generally found in the training data of AI systems.

But when people with this domain expertise design products to address the problem that they're familiar with, they can bake this expertise in.

They can do this by:

• Fine-tuning AI models to the domain.
• Defining step by step workflows for the AI system to follow.
• Providing a better user interface, specialized to the task at hand.

Conclusion

As agents advance, they will unlock more and more value:

• Cheaper labour: Layoff opportunities, increased profit margins.
• Tireless workers: Supply bottlenecks broken and demand satisfied.

Like self-driving cars that were the hot topic in 2016, but are only just rolling out in now, 2025 likely marks the dawn of the decade of agents rather than the year.

Until then, agents still need operators before they can take on the full role of a teammate.

Hype:

• Websites that claim to have a suite of different agents for doing anything - there's always a specialized, focused alternative that will outperform.
• Agents that can run your entire company.
• Agents can't completely replace team.

Real:

• Agents can massively accelerate many different types of work.
• Where they exist, use-case specific agents are capable of completely automating job functions, thanks to their team's focus.
• General purpose agents are real, but claims are often overstated.