A.I. Agents With Agency: The Inflection Point

It’s no secret that I’m obsessed with agents.

But that obsession actually began over two years ago (way back in 2023!). This led me to talk about Agent AI on stage at INBOUND 2023 in front of 10,000 people, where I defined agents simply as "software that uses AI to accomplish a multi-step goal."

Today, I want to share how my thinking on agent agency has evolved, where we're headed with agent.ai, and why this matters for anyone using or building AI agents.

In this post, I'll break down:

• The spectrum of "agenticness" and why it matters

• Why we started with simpler workflow agents

• The reasoning revolution that's changing everything

• How tools dramatically expand what agents can accomplish

• Finding the right balance between autonomy and predictability

—@dharmesh

The Agenticness Spectrum

Back in 2023 at INBOUND, I defined agents simply as "software that uses AI to accomplish a multi-step goal" (you can watch the replay here, if interested).

I deliberately avoided requiring agents to be autonomous or have high "agency" in my definition because I didn't want to wait for perfect technology to start building useful things. Even deterministic AI could be considered an agent if it accomplished multi-step goals.

I've always thought of "agenticness" as a spectrum:

• On one end, you have deterministic agents following predefined steps (high predictability)

• On the other end, you have autonomous agents making their own decisions (high agency)

• Most useful agents fall somewhere in between

When we started building agent.ai, we deliberately focused on the deterministic side of this spectrum — what some call "workflow agents" (or AI workflows).

The reason was simple: in 2023, LLMs couldn't reliably break complex goals into logical tasks, develop execution plans, and handle unexpected situations.

As agenticness increased, predictability decreased dramatically:

This graph illustrates why we prioritized predictability early on. By focusing on workflow agents with clear, defined steps, we could deliver reliable results while the technology matured.

I'm happy with that early choice because it helped us build something real. Agent.ai now has 1.6M users, with 17,000 builders creating agents of varying sophistication.

But something important has changed in the AI landscape recently...

The Reasoning Revolution

The moment we've been waiting for has finally arrived.

Thanks to new models like OpenAI's o3, LLMs now have pretty strong reasoning capabilities, and it's actually possible to increase the agenticness of an agent and still get reasonable predictability towards the desired outcome.

This is what the updated graph looks like in 2025:

Note: These are not drawn to any kind of scale and are not based on data. They’re just a way to illustrate a point.

So, I've been working on allowing the agent.ai Agent Builder to create agents that can reason.

You describe the goal, give it access to a set of tools, and then let the LLM (usually OpenAI’s o3) do most of the work.

For example, instead of having to define every step in a workflow, you can now say "Review the recent tweets of @yaminirangan and email me the 3 with the most engagement." and the LLM can figure out:

1. That you’re talking about posts to X/twitter

2. That there’s a tool in agent.ai available to get recent tweets

3. Get the recent tweets and find the ones with most likes and replies.

4. Use a tool to email me the results.

The agent can now reason about the problem rather than just following a script. This dramatically expands what's possible.

The Crucial Role of Tools

Tools are super-important in making these more agentic systems work well (this is worth its own post soon as well).

They help LLMs do much more than their training data would allow, such as:

• Allowing an LLM to access third-party data sources

• Taking specific actions in external systems

• Searching the web for up-to-date information

• Processing and analyzing data in specialized ways

• And so much more…

Part of what has me so excited about Agent AI is that every agent on the network will effectively become an available "tool" for other agents to use (I wrote more on Multi-Agent systems in a previous post).

This has been the dream all along — and we're getting closer and closer to realizing it.

Imagine a knowledge management agent that can leverage a scheduling agent, which in turn uses a data analysis agent, all working together to solve a problem no single agent could handle alone.

The possibilities are endless when agents can collaborate through a standardized protocol.

The other advantage of a multi-agent system is that because each agent is “specialized” and trained/tested to do one specific task, predictability goes up. It’s classic software engineering — break the problem up into pieces such that each piece is easier to solve.

Finding the Right Balance

So, what does this all mean for how we're building agents going forward?

Well, I don't expect us to abandon workflow agents completely. You should only rely on LLM reasoning if you really need to.

For example, if you know what steps need to be executed to get a job done, then you're likely better off just expressing that in a deterministic workflow. You'll be happier for it.

The right approach is finding the optimal balance between deterministic workflows and autonomous reasoning for each specific use case. At a high level:

• For repetitive, well-understood tasks with little variation → Workflow agents

• For complex, dynamic tasks requiring adaptation → Reasoning agents

• For many real-world applications → A hybrid approach

At agent.ai, we're now extending our platform to support both approaches seamlessly, allowing developers to apply the right level of agency for each specific need.

We're at an inflection point in agent evolution. For the first time, we can create agents with meaningful autonomy that still reliably accomplish their goals.

Agents with just the right amount of agency are going to transform how we interact with technology. And the network effects of connected, specialized agents working together will create capabilities far beyond what we can imagine today.

That's why I'm building agent.ai, and why I’m more excited than ever about what comes next.

—Dharmesh (@dharmesh)