A first-principles framework for agent–human collaboration across research, decision, and monitoring.
This paper sets out a first-principles view of how AI reshapes fundamental investing. Genuinely AI-native systems — redesigned around agent–human collaboration rather than bolted onto legacy workflows — create durable advantage specifically in fundamental investing beyond three months, where the game is probabilistic prediction over vast bodies of information.
It examines what agents and humans are each good at, how work should be divided, the prerequisites for AI-native teams to match and then surpass human teams, where the technology stands today, and how a legacy fund makes the transition.
Once the constraints that bound human investing fall away, the optimal move is to run the most rigorous analysis on as many ideas, as often, as possible.
Move along the holding-period spectrum. The durable edge sits at the far right.
Why is the long-horizon game different? No single data feed, proprietary call, or trend line deterministically predicts an outcome past the quarterly cycle. The work is large-scale information synthesis and pattern recognition — exactly where AI beats humans. And the market won't shrink as AI arrives; it grows more vibrant, because three properties hold.
No one — human or AI — predicts perfectly. Many winners coexist, and any system can keep improving.
Returns trend positive over time as society advances — which is why new GPs launch every year.
Different horizons, markets, and sectors each demand their own framework to predict well.
How many companies are considered.
The quantity and quality of info synthesized.
The correctness of analysis applied to it.
The framework used to reach judgment.
The personal capital that drives fundraising.
An agent leaning on raw model knowledge inherits the average of the internet — and cannot, in principle, out-judge the best human investors.
Not everything in a great investor's head is written down, and what is gets polished. What worked for one investor won't transfer wholesale to another.
Analyses, frameworks, and strategy must cohere. Horizon, sector, market, and the investor's own strengths all shape the right guardrails.
Agents can refine on top of a framework — but they need it guided and encoded by humans. And whoever stays accountable for the system must understand its guardrails deeply.
As digitization improves, everything on the internet should be considered accessible. But the internet doesn't reflect reality — emotions are performed, propaganda and censorship abound, some voices dominate, and most people stay silent.
Skilled investigators surface what people won't volunteer online. Economic temperature is felt on a busy street, not in cold reporting. Seeing and interacting carry higher resolution.
The exception: longer-horizon strategies on stable businesses lean on established online facts — and a basket approach dilutes any single offline edge.
Tap each requirement
These are the conditions an AI-native team must meet just to match a strong human team — before it can surpass one.
Tech-stack maturity extrapolates toward feasibility; human behavior and system iteration may delay the timing of real success.
On a clear path toward inflection points, depending on model and harness.
Design patterns are maturing, but seasoned investors lack the time, mindset, and setup to transfer what they know.
Clear pathway — meeting bots, expert transcripts, RAG-enabled stores — not yet perfected.
DIY builds underestimate the burden; legacy tools only assist existing workflows rather than enabling AI-native practice.
Trust in agents fluctuates; the shift is hardest for older, successful organizations.
Illustrative figures. The principle: do one more analysis whenever its marginal value beats its (falling) token cost.
Death to 80/20. Burn the tokens, and do all of it.
High-quality sources, as broad as possible.
Surfaces the right information and applies the right skills without errors.
Ideation, macro, qual & quant analysis, private-info integration, human review, portfolio construction, risk, monitoring, buy/sell.
Skills, multi-step orchestration, and scheduling — all customizable with little effort.
Human rules of thumb, agentic reflection on past trades, reusable wisdom, and a conflict-resolution agent to keep it coherent.
Relative return over a benchmark.
Sharpe and similar measures.
Performance through regimes.
Long feedback cycles and many interacting variables make attribution hard — but that's no different from how human investors have always been judged. Frame the fund as an evolving system: every component improves over time, and performance should follow.
PMs own the relationships, the strategy, and the team — so the decision is theirs. Once committed, the hard part is steering the chaotic in-between without breaking risk controls or culture.
Map the current strategy → framework → analyses, then work backward to find the analyses agents can perform.
Assign those analyses to analysts; have them encode the guidance and steps into skills and multi-step flows, and own the results.
Combine the tested blocks into partial automations — screening & ideation, initial analysis, monitoring, buy/sell, portfolio construction.
Operate each automation beside existing routines as a check and supplement; refine frameworks toward optimal output.
Move analysts from hands-on analysis to monitoring the automation — and to sourcing proprietary information offline.
Run the full cycle without scaling yet, with analysts checking and refining the system.
Keep improving the system and developing new routines.
Scale ideation breadth, analytical depth, and frequency until performance surpasses human teams.