AgentBull
Intelligence amplifiers for consequential decisions.
AgentBull is a Neolab that builds intelligence amplifiers for key decision-makers through outcome-oriented AI agents. We are paid for the business outcomes our agents achieve, and we put what we build before the most honest examiner there is: real markets, with our own capital at stake.
Four terms
- Outcome-oriented
- We do not charge per seat or per call. We charge for the business outcome actually achieved—the unnecessary inventory a client no longer carries, for example—and take a share of that value.
- Key decision-makers
- Our trade-off is to concentrate on decisions assisted by AI rather than on execution that is merely mechanical. So our users are people who carry independent responsibility for a decision and whose own interests ride on its result: executives, fund managers, ad buyers, salespeople.
- Intelligence amplifier
- The phrase comes from Douglas Engelbart's 1962 paper Augmenting Human Intellect, whose central claim is that the greatest value of a computing machine is to make human judgment stronger. Our AI does not decide for the decision-maker. It lets them see further, consider more, and err less.
- Neolab
- We are a research startup, not an application company. We study the state of the art and build both our moat and our revenue on it: push the impossible to possible, then turn the possible into commercial value.
Product lines
Ankole
Open source · Free
Ankole is an open-source, enterprise-grade Agent Harness: a harness that gives a company a brain and genuinely assists its decisions. Enterprise-grade means its agents are colleagues inside the company rather than personal assistants. Each has its own mission, KPIs, and permissions, and can work below, alongside, or above human staff. The company brain is a living map of how the company actually runs. The knowledge a company values most rarely sits in documents; it sits in the heads of long-serving employees, in chat histories, and in dozens of internal systems that do not talk to each other. Ankole extracts those fragments, works out how they relate, and keeps them current as the business changes. It is not a knowledge base that answers questions. It is an enterprise world model, kept up to date.
AI Associate
Commercial · Private deployment
AI Associate is the commercial edition of Ankole. It provides higher-quality decision support and, through deep inference, a sense of where things are heading, optimized for settings such as capital markets where the bar for decision quality is higher. Its Deep Research is built for questions of the form “how likely is this, and which way will it go”: the odds of a crisis in a region, how far a stock will move, how a product will sell, most of the events listed on Polymarket. That is a different task from the Deep Research in general-purpose assistants, which organizes what is already known into a report. Ours reasons about what has not happened yet.
BullXCloud
Cloud platform · API / MCP
BullXCloud opens our strongest capabilities and our most valuable data to agents, through APIs and MCP, for Ankole and AI Associate to call. The capabilities are our own models, a search built for agents rather than people, and Deep Research Ultra, the flagship tier, offered only through the API. The data is intelligence-grade signals built from open-source intelligence, news, and market data, up to signals a decision can act on directly, such as forecasts of near-term direction and volatility for individual A-shares and industry sectors.
Why a harness
The stronger the base model, the more a harness is worth.
A base model supplies general capability. But however far it evolves, its training remains statistics over historical data: it tends toward the most common answer, not the answer that is right for this specific goal, now.
A high-quality decision depends on an accurate projection of the world as it is at this moment. A model's knowledge has a cutoff date, and search and knowledge bases patch in fragments without ever assembling the whole. A real decision also turns on things no model can ever learn.
What no model can learn in advance
- 01The latest state of the real world
- 02Context inside the organization
- 03Tacit knowledge that was never written down
- 04The decision-maker's own goals and preferences
- 05The permissions and boundaries of action
- 06The real feedback after each action
These exist only in the moment the work is being done. They happen at runtime, and they cannot be trained statically into model weights. That is where the harness sits. It is not glue between a model and its tools. It is the runtime system that governs the AI's attention and behavior: its judgment, and its self-restraint.
It decides what deserves attention and what can be ignored; how much is enough and when to stop; which judgments and corrections to keep for the long term and which apply only to the task at hand; and how the result of an action should change the next decision.
A base model is raw material, and a stronger model is better raw material. The system that turns raw material into high-quality decisions is us.
A system, not a blueprint
Taxonomy tidies drawers. Structure stacks bricks. Dynamics grows an ecosystem.
Agent harnesses have so far evolved only three worldviews. The first classifies: planning, memory, tools, and execution each get a box, drawn from the top down and then filled in, like an organization chart. The second composes: the system is broken into atomic bricks that can be assembled any way you like, like an anatomical drawing. Both deliver a blueprint, fixed the moment it is drawn.
Ankole takes the third. It asks neither what a system is nor how it is assembled, but how it is running now and where it goes next. Behind that stand two disciplines with decades of history, dynamical systems and cybernetics. Their shared lesson is that a system reaches its goal in a changing environment not by following a perfect plan but by sensing, acting, observing the result, and correcting, the way a helmsman corrects the rudder every second.
In practice, Ankole treats a system as a population of long-lived objects, each holding its own state, cooperating through signals. Every agent is like a living employee: it has its own state and memory, it acts when a signal arrives, and the result of its action becomes a new signal flowing back into the system. Such a system is not assembled; it runs. It is not the company's organization chart. It is the company, running.
Why the difference is not rhetoric
- 01
The unit is a signal, not a chat
A chat is one kind of signal. An order entering the CRM, a stock moving, an event thrown by an external system are signals too. An Ankole agent does not wait for someone to open a conversation. A blueprint exists only while someone is looking at it; a dynamical system is always running.
- 02
Feedback acts on memory too
Every night, Ankole consolidates fragments of memory into higher-level regularities, lets what no longer matters cool and lose weight, and grows the company brain's ontology from the bottom up instead of designing it in advance, as Palantir does. Forgetting is itself part of intelligence: when everything keeps equal weight forever, what truly matters drowns.
- 03
Homeostasis is where the hard capabilities come from
A body heals without anyone redrawing its anatomy. Tasks that resume from where they stopped after a power cut, processes that recover from a crash, thousands of agents cooperating at once: all follow directly from one principle, rather than being built one feature at a time. No anatomical drawing can depict being alive.
The first two deliver a drawing, fixed the moment it is finished. Ankole delivers a system that breathes and corrects itself.