Becoming Agent Bill Ackman
Maintaining Context for Decision Making in Financial Supermind Agents (SuperBill)
Why did the stock drop?
Is actually an interesting question.
Not because of your FOMO rising. Mainly because it refers to an event that has happened in the past. Therefore, theoretically all information required to answer it should be available. If you only would know where to look. A typical needle in the haystack problem. But it must be there. Somewhere out there in the vast digital realm of the Internet. If one would just manage to have perfect information about it at their fingertips. But there isn’t. Yet, this is a herculean task that Bloomberg and Reuters build multi-billion-dollar businesses off from.
If only it would be easy to answer that question it would be less value.
Figuring out the past is one thing. How about the future?
Will the stock drop?
Also an interesting question. But here, not all information is available because the future has not happened yet. All future outcomes are still possible.
All of them.
Stocks fluctuate for a variety of reasons. People want to cash in before tax season, rumors of a cheaper model coming out of China (DeepSeek), political changes, unsuccessful product launches (EA), or some of the many other reason (NVIDIA) that make the capital markets such a fascinating realm to work in.
Even legendary investors like Bill Ackman, analyze trading patterns, deal flows, news sentiment and many other indicators to predict what the value of an equity will be in the near to mid-term future. Some call them swing traders, momentum, or day traders. There are many investors who think they have found the secret to an infinite money glitch. And some, like Bill Ackman, do.
So when building SuperBill — an AI Agent for investing, what do we need to do?
Well, what all of these traders have in common is that the act of trading is will always be seen as an “action in context”. Hence, I’d argue that if for us meatbags context is important when making investment decisions, then context is also important when building SuperBill.
But what is context?
Context Discovery
Context in language refers to the background information, circumstances, or shared knowledge that influences the meaning of words, phrases, or sentences. It helps disambiguate meaning and provides deeper understanding.
Context can be:
Linguistic: The words surrounding a phrase that affect its meaning.
Situational: The external circumstances in which communication occurs.
Cultural: Shared norms, history, and beliefs that shape interpretation. Japan
For example, the phrase "That's a big deal."
In an earnings call, it may mean an important contract has been signed.
In casual Reddit commentary, it might just mean something surprising maybe significant. But just “big deal” could be even derogative or even indicate a negative connotation.
For decision-making in an investing context, investors should gather as much information as possible about a trade, as context includes the factors that influence and shape their choices. Examples for such priors are any constraints, goals, and external conditions. Clearly, having good data (accurate and timely) is expensive.
A fantastic resource here is Bill Ackman as he talks about value investing.
A decision taken in wrong context often leads to suboptimal outcomes.
Context-aware decisions would consider — and is clearly not limited to:
Historical Data: Past trends, experiences, and learned patterns. LTSM might be the “brain” you need here.
Environmental Factors: Market conditions, regulations, and competitive landscape.
Stakeholder Context: How different parties will be affected — who are stakeholders anyways?
Macroeconomic Context: Interest rates, inflation, and geopolitical risks.
Industry Context: Whether the company’s sector is growing or declining.
Regulatory Context: Any new government policies affecting the company.
Historical Context: The company's past performance during similar market conditions.
So, you want to buy NVIDIA?
Well, what would Bill do? The stock “just” dropped by 20% – …well “just” depends on when you are reading this. Time is also a measure of context. So, is now a good time to buy? Imagine you a fund manager deciding whether to invest in 10 million USD or not. Without the right context, they would look at the company’s recent profits and product announcements and decide it’s a good investment.
In the future we might have agents that have been trained with all the experience, knowledge, and context of Bill Ackman.
For Agents though, context is more complex.
And in most of these cases, what we take granted as humans, we specifically need to define for agents. What is the data we get, how will be encode/embed it so we have a shared context space across agents and interactions.
But also,
what are the tools to manage them,
what are measures to evaluate them, and
how can we integrate them into our analysis
As we are building SuperBill, we need to manage the context he will operate in because this context represents the world model SuperBill grounds its truth in.
Allow me to go on a short detour. When I was doing the exercise of a 4x4 board game, the state of the board game together with the list of available positions on said board laid out a clear world-model for the agent to operate in. This ensured that the agent’s response could me tracked, categorized, and measure. Because of that, it was noticeable that the reasoning process of the agent still failed in most cases.
Why was that?
Context Management
In this example, context is working memory. When developing the game use-case, I failed to implement several context management activities throughout the engagement correctly.
Task continuity: During agentic workflows that involve multiple steps or interactions, the correct maintenance of task continuity ensures that on each step of the way, we are aware of previous actions, decisions, and data, guiding us to reliable and consistent outcomes.
Error Repetition: Agents have a higher likelihood to repeat the same error, never terminate thus wasting resources and time. We also need to ensure that the information written into context state is not repeated.
Accuracy. In all statistical models errors exist. In basic linear algebra models, we used to track them with MSE, RSQ or similar. Now measuring the quality of a generated body of text is really hard. Maybe even impossible if you want to govern your agent flows on scale. Therefore, it is ultimately important to define extremely narrow guardrails through the context and task description to the agent and measure the output through this concept. I.e., ensure that the agent selects a valid move in the game or selects a valid price point of a stock. What is the gap between expectation and real result.
Resilience: In case of errors or interruptions, context allows agents to recover more effectively by understanding what has already been done and what needs to be redone or corrected. For example, if the agent is executing a trade on Monday and a trade on Friday, then the information from Monday has to be in context.
Audit Trails: Maintaining context helps in creating audit trails, likely through an additional tool call. This is essential for compliance with internal and external regulations and standards.
Accountability: Context accountability ensures that each action taken by an agent can be traced back to a specific decision or input, providing accountability. For having a reliable system you want to be in a position where you can quickly judge who the culprit agent is,
Handling Complex Tasks: With more agents in the workflow and higher structured tasks, complexity to manage context correctly increases. To succeed here, it is essential to break down the task into manageable parts and ensuring that each part is executed correctly.
Dynamic agent architectures.
As workflows scale, the complexity increases. We need to have a clear understanding of the dependency graph in the agent network. However, we also learned already that we should expect that data flow through agent networks change dynamically as well. Even though that might seem to contradict some of the earlier points raised.
How would you manage context in this case? For example, how do you ensure that the graph adjust correctly if you add a new data source, i.e., like Seeking Alpha’s Earnings call.
In closing - Multi-Agent Systems
In workflows involving multiple agents, context is essential for coordination and collaboration. Each agent needs to be aware of what others have done to avoid conflicts and ensure smooth operation. When humans and agents work together, maintaining context ensures that the agent understands the human's intentions and can provide relevant assistance. Moreover, over time, agents can learn from past interactions if context is maintained, leading to continuous improvement in performance. But it’s likely easier if the agent stays stable, especially if you consider model, and also improves on new in-context information.
Establishing and maintaining context in agentic workflows enables agents to handle complex tasks, adapt to changing conditions, collaborate effectively, and scale operations while maintaining compliance and accountability.
Without context, agentic workflows would be prone to errors, inefficiencies, and a lack of coherence, ultimately undermining their effectiveness.
And no, I am not advertising $BABA. But their Qwen model is really good.
