Where Are We on the Bot-Agent Spectrum?

Lately, the word “agent” seems to be everywhere, conferences, new projects, social media. Everyone’s talking “agentic,” though I’m not sure there’s a perfect Turkish translation. I’ve mentioned AI agents a few times in past articles, talking about their different uses.

But honestly, I’m not sure how accurately we’re using the word “agent.” Is everything we’re calling an agent really an agent?

Last year, Gartner gave this phenomenon a name: “agent washing.”

Basically, it’s relabeling existing chatbots and rule-based automation tools as “agents.” Many projects and producers are making this claim. According to Gartner, only about 130 truly function as agents.

The same report predicts that over 40% of current AI agent projects will be canceled by 2027, due to rising costs, unclear business value, and inadequate risk controls.

So here’s a question worth asking: What exactly is a bot, and what is an agent? What’s the difference?

We all have some idea of what a bot is. Bots have clear roles, whether as chatbots or in cybersecurity, and in finance, they have a very specific use: they’re automation mechanisms that follow a set of rules to trade investments.

There are many indicators used in trading investments. These indicators are mathematical formulas used for buy or sell signals.

Bots execute buy and sell actions based on signals from various indicators.

For example, RSI (Relative Strength Index) is a technical indicator that measures the speed and magnitude of price movements. You can set a rule in a bot to buy when RSI falls below 30 and sell when it rises above 70. This rule doesn’t change, the bot doesn’t learn from its trades, and it doesn’t evolve its rules. No matter how market conditions shift, the rule stays the same. Adding more indicators or increasing parameters doesn’t turn a bot into an agent. It just makes it a more complex bot.

On the other end of the trading automation spectrum, there are systems like JPMorgan’s LOXM, announced in 2017. LOXM uses deep reinforcement learning, meaning it learns through trial and error from billions of simulated and real trades. It’s designed to manage large orders: how long to execute them, at what price, and in how many pieces. The system makes these decisions based on market conditions and learns from the results of its actions. The model’s weights, the decision-making process within the system, are constantly changing and evolving. This is how true learning works in the AI world.

Back in 2017, this technology was led by major institutions and accessible to very few individuals.

Today, things are quite different, and this is significant because the real story starts somewhere between these two types of automation.

It’s not just a bot with a set of rules, but not quite a fully learning model like LOXM either. How does this “adaptive strategy” work? You set a few different parameters on the indicators we mentioned earlier to open a trading position.

These parameters essentially become your strategy. Trades are opened based on this, and results are recorded. The strategy is periodically updated based on these results. For example, if the threshold set for RSI isn’t working, it’s changed.

A memory layer is added to the system: a vector database, a structure that stores past trade results. Even if the computer or session is closed, the past remains there, and the system uses it in the next trade. The model doesn’t change, but the memory grows and the strategy updates.

Until very recently, implementing this as an application was much harder. But today, this architecture can be set up with exchange-provided MCP connections, open-source reinforcement learning libraries, and locally run large language models.

The gap between what institutions took years to build and what a developer can set up in a short time is narrowing.

It’s a system we can experience firsthand through different exchanges’ MCP servers. I’ve mentioned before that OKX has opened up its infrastructure, but I’m sure others have it too. The technical infrastructure for agents that can update themselves, even if they don’t learn, is now accessible, but the line between a “truly learning system” and a “well-designed bot” remains thin and important.

Beyond the technical infrastructure, theories about where to buy and sell investment tools and price movements, which are entirely based on human psychology, only apply in a world where trading was done solely by humans.

In recent years, bots have already changed this. They’ve shaken markets from time to time. In 2016, the British pound’s overnight 6% drop turned into a market movement triggered by algorithms. By the time it was noticed, the price had already changed.

So if bots with just rule sets can impact the market this much, how will the dynamics change when systems that learn, adapt, and have memory become more widespread? And are the markets truly ready for this? Are we ready for it?