Quantitative libraries used to be written once and then defended. With AI they can be rewritten continuously — but only if they are built for it.
The static model
Today the infrastructure is fixed, the libraries are fixed, and the work consists of making them run together. The revenue engine — trading in production — cannot stop, while infrastructure and software are counted as cost. Any structural change is expensive and, worse, disruptive — so patches and workarounds win over the right fix. Each patch makes the next change harder, until any development that touches the infrastructure becomes a Herculean task. This holds within a single library and across a stack of libraries.
The same system also resists automation, even with AI. An AI tool dropped into it meets undocumented conventions, inconsistent code and hidden dependencies — the obstacles a new hire meets, with less context.
The dynamic model
In the new paradigm, libraries are AI compatible: an AI can read them, understand them and change them safely. An infrastructure built from them is no longer fixed. It can be updated at will, adapting to new realities — new products, new regulation, new data — as they emerge, instead of being bent around them.
Four things, which only work together:
- Clean code. Usually written by the Quantaiko AI Agent — not mandatory, but code written by hand tends to be inconsistent and rarely bug‑free.
- Documentation written for AI. Human documentation aims to be exhaustive; AI documentation aims to be relevant. A model works from what it has in context: give it too much and it draws on the wrong material, taking a path far from the best solution; give it too little and it has nothing to draw on. AI documentation is therefore layered and targeted, so each task retrieves exactly what it needs.
- Code that matches its documentation. Documentation is useless if the code does not follow it precisely. The code keeps one consistent style, does exactly what its documentation says, and stays easy to read — for an AI and for a human. Keeping the two in step is itself a strong incentive to have the Agent make the change rather than code it by hand.
- A set of external tools provided by the Quantaiko AI agents. These are indispensable: AI makes mistakes and is slow, whereas these tools give correct, fast results for specific tasks.
- Change becomes routine. Interface and infrastructure changes are fast to implement, rolled out in stages and tested systematically at each one.
- Fewer bugs. The mundane coding is handled by AI under the supervision of the specialized AI Agent, with checks at every level; tests are written and run automatically with coverage tracked at every step, and debugging is carried out by the agent.
- Operational risk is better understood and mitigated, because every change is small, tested and documented.
- Support and new development are straightforward. The purpose‑built documentation produced by the specialized AI Agent lets an agent “understand” the library and make the right connections across it, so support is far faster and is never blocked by unfamiliarity with the codebase.
- Model validation and compliance become automatable. The agent produces the tests and the validation documentation, and any newly required tests and documentation are delivered in hours.
- Full library optimization.
- New ideas reach production much faster — a much shorter cycle from development to production.
- Language agnostic (C++, Java, C#, Python, …).
One of the most significant consequences of our dynamic approach, see workflows, is that libraries can now be optimized to a degree that was previously out of reach. Established practice pushes towards code that is as generic as possible, so that it can be reused across projects with minimal — ideally no — modification. The price of that generality is optimization: a generic library can never be tuned as far as we would like. The obstacle was never confined to specialization over primitive types (integer, floating‑point, and the rest); it sat at the infrastructure level, where full optimization would have meant rewriting the library outright.
That constraint no longer applies. The agent makes fully optimized libraries practical to build, and Time Series Predictor is the clearest illustration: we produced several specialized, optimized libraries — specialized LLM libraries among them — each shaped precisely around the way it is used. The gains have been substantial.
Flawed ideas about the mathematics or the infrastructure lead to flawed foundations, and everything AI builds on top of them inherits the flaw. The result is not merely imperfect; it is worthless.
This is the trap that is too rarely acknowledged. AI lets you travel much faster. As long as you work within your own expertise, you can steer it onto the right path. Beyond that expertise, AI remains just as fast, but you no longer know where you are going and can no longer guide it. It will then take you somewhere you never intended to go, and you are unlikely to notice until it is too late.
The use of AI also differs markedly from anything taught in our studies or practised on the job. It opens access to an enormous body of knowledge, both sound and unsound, and drawing on it reliably demands a new discipline: knowing how to frame questions, and how to judge and categorize the answers. AI is still in its infancy, and today that judgment is learned mainly through experience.
For these reasons, the work cannot be handed to inexpensive junior staff. It calls for people who combine finance, deep software engineering, and genuine AI experience. We built the Quantaiko AI Library Agent precisely to orchestrate other agents, such as Claude Code, and to put the full power of AI at your company’s disposal, guided by experienced quant, engineering, and infrastructure specialists.
The gains from working with the Quantaiko AI Agent are not uniform: they differ by the phase of work. Measured against experienced manual coding, our estimates are:
- Research and design — a factor of 2 to 5.
- Implementation — a factor of 5 to 50.
- Support — a factor of 5 to 10.
- Infrastructure changes — a factor of 10 to 100.
Research and design — a factor of 2 to 5
This is where you need experienced people, in the domain — banking, in our case — and in infrastructure. It is the foundation of the library and of how the library is organized, and wrong foundations lead to catastrophic results. AI gives you a vast database of what already exists, ready to answer your questions; but if you do not know what to ask, or how to ask it, the answers will not help — and they may be wrong, even harmful to the later phases. In experienced hands, by contrast, AI deepens what the team already knows and helps it reach robust solutions.
Implementation — a factor of 5 to 50
This is where automation delivers the most, and where the Quantaiko AI Agent can push hardest. The value depends on the foundations: if they are weak — a legacy library, or a first phase that was not done well — automation only produces the same flawed work faster. And good foundations alone are not enough: even knowledgeable people, using AI on its own, can write poor code. This is one of the most common traps, and it holds even with an agent such as Claude Code. Used directly, without an agent, AI might earn a knowledgeable developer a factor of about 2 — but checking the output is painful. With an agent, you must still know how to work with it, and much remains manual: checking the code, writing the tests, understanding the errors. Here the gain is perhaps a factor of 3 over manual coding.
The Quantaiko AI Agent lifts this to another level. From its generic AI‑agent library it first builds your specialized AI Agent adapted to your own, so that it can converse with agents such as Claude Code, and it carries out many of the remaining steps itself with its extensive set of tools. That adapted library keeps growing and fits your library better every time it is used. Code creation becomes automated to the point where the experienced person spends ten or twenty minutes scripting the plan, and the Quantaiko AI Agent then works on its own — for several days if need be — to produce code that is:
- certified to a high standard and adapted to your infrastructure;
- verified with a very extensive suite of tests;
- AI compatible, with AI documentation;
- accompanied by human documentation.
Depending on the library, the gains are substantial, easily reaching a multiple of tens. Taken together, the first and second phases — a new implementation — bring an end‑to‑end gain of 4 to 40 compared with manual coding.
Support — a factor of 5 to 10
Support is another area where the Quantaiko AI Agent shows its potential, provided the first two phases were carried out with the agent and the library is AI compatible. It covers fixing bugs, improving models, and adding models or products. The agent has excellent AI debugging tools: combined with the AI documentation, it locates the cause of a bug and fixes it quickly. Improving a model — adding parameters and code — is straightforward, because from the AI documentation and the code itself the agent knows what to change and creates standardized new tests. Adding a model or a product benefits in the same way: a task that could be painful in a legacy library — and we met several such cases in our careers — becomes far simpler from the developer’s point of view, because the difficulties are handled by the agent. The support of an AI-compatible library therefore consumes far less developer time, by a factor of 5 to 10.
Infrastructure changes — a factor of 10 to 100
This is where the Quantaiko AI Agent excels. Introducing infrastructure changes into a library that is not AI compatible is hard, to say the least; across several libraries it becomes a nightmare. You never know whether you will introduce a downstream bug, because a library is not used in the way intended, so you tend not to update it — and you end up with an infrastructure full of patches, impossible to change. With AI‑compatible libraries this does not happen: the agent knows enough to change the interfaces — not everything, but enough, and that is the key. It understands where a change could cause problems and creates extensive tests to cover them. What you would normally dread, and what usually takes groups of developers years, becomes a systematic task that the agent handles gracefully. This is why the gains are so large: tens for localized changes, hundreds for global ones.
The agent combines an extensive Claude CLI prompting library with a toolkit of proprietary and external tools to drive dynamic planning‑and‑execution workflows. The prompting library has been calibrated to deliver an optimal result.
User input remains fundamental — but it is concentrated on the flow infrastructure, not on the coding itself. This is the clearest departure from the old‑school developer role.
The agent serves a broad range of engagements — see the Applications it has produced:
- Re‑implementing large existing quantitative libraries. We have drawn on our years of experience as front‑office quants at major investment banks to optimize the agent for rewriting these libraries into clean, well‑formed, thoroughly tested, and extensively documented code, while preserving their original interface, making adoption seamless and reducing support overhead. As a result, valuable resources no longer need to be spent supporting legacy code, and can be reallocated to more revenue‑generating projects. NumpyPandas C++ is one such re‑implementation.
- Research. Automation lets a hypothesis be formulated, tested, and kept or discarded at a speed no manual process can match. Time Series Predictor is one example, and the Optimal Dynamical Portfolio built on top of it is another — turning those predictions into a portfolio that hedges, optimizes a market view, and raises the Sharpe ratio; see also Optimization.
- Building to specification. The client sets the objective and the specification; the agent delivers against them, and delivers fast. The CUDA CPU Simulator is one example — a comparatively straightforward case, since existing simulators, over‑complicated though they were, already defined the target.