Andrew Ang, Columbia University adjunct professor and former head of factor investing at BlackRock, published a paper in early April describing a system where approximately 50 specialized AI agents autonomously execute the full strategic asset allocation process. The paper, titled “The Self-Driving Portfolio,” was co-authored with Nazym Azimbayev, a sovereign wealth fund CIO, and Andrey Kim, a Deutsche Bank quant, according to AI Street.
How the Pipeline Works
The system is governed by the Investment Policy Statement, the same document that constrains human portfolio managers. Every agent reads it. A chief risk officer agent checks compliance for every portfolio candidate, per the paper.
The pipeline starts with a macro agent that classifies the current economic regime using data, market indicators, and web searches. Asset class agents then run in parallel, one per asset class, each estimating expected returns using six different methods before blending them into a composite. An LLM-as-judge step reads all seven estimates alongside the macro regime and selects a final figure with explicit weights and written rationale.
Twenty portfolio construction agents each build a portfolio using a different method. A 21st researcher agent scans academic literature and proposes methods not yet in the pipeline. A separate adversarial diversifier deliberately constructs the portfolio most different from consensus, as detailed by AI Street.
Agents then peer-review each other: each reviews two others (one similar, one different), vote on rankings, and flag bottom picks. A CIO agent combines top candidates using seven aggregation methods and produces a board memo for non-technical stakeholders. A meta-agent compares past forecasts against realized returns after each cycle and rewrites other agents’ code and prompts to improve performance.
The March 2026 Test Run
The authors ran the pipeline against a mandate covering 18 liquid asset classes (6 equity, 8 fixed income, 4 alternatives), targeting CPI +3 to 4% real return with 8 to 12% volatility and a maximum drawdown of negative 25%.
The macro agent classified the current environment as late-cycle with stagflationary risk. Asset class agents responded by discounting historically expensive markets: US Growth stocks had forecasts cut 2.0 percentage points below composite, US Large Cap cut 1.1 points, while Emerging Markets were barely adjusted, according to AI Street.
The agents collectively favored methods relying on historical volatility and correlation data over return predictions. Maximum Diversification ranked first. The final portfolio came out at 44.9% stocks (vs. 60% in a standard balanced portfolio), 41.7% bonds, and 8.1% cash. Backtested from 1996 to 2026, it produced nearly the same return profile as a 60/40 portfolio with a peak-to-trough loss of 25.6% versus 34.3%.
Proof of Concept, Not Strategy
The authors frame this as a proof of concept. One run producing a sensible portfolio over a single test period does not validate the approach for live capital deployment.
The paper’s framing is institutional: the bottleneck in asset allocation is bandwidth, not analytics. A CIO can supervise 10 to 15 departments. A research team can cover 20 to 30 asset classes before the process stalls. The argument is that agent swarms can compress a quarterly investment committee cycle into a documented, auditable pipeline that runs on demand.
For teams building agent architectures outside finance, the structural patterns are transferable: parallel specialist agents feeding into peer review, adversarial diversification, meta-agents that rewrite downstream prompts based on realized outcomes, and governance documents (the IPS) constraining autonomous behavior at every step.