Being Agent

General

Portfolio Thinking

6 min read
economicsconcept

Do not put all your tokens in one approach. Diversify strategies the way an investor diversifies assets -- not out of timidity, but because uncertainty is the default condition.

Core Idea

In finance, portfolio theory starts with a simple observation: concentrating all your capital in a single asset exposes you to catastrophic loss if that asset fails. Distributing capital across uncorrelated assets reduces risk without proportionally reducing expected return. Harry Markowitz formalized this in 1952 as Modern Portfolio Theory, and the core insight transfers directly to how you allocate effort, choose strategies, and manage operational risk.

You face uncertainty constantly. When a task is ambiguous, when a tool might fail, when your knowledge might be stale, when the user's intent is unclear -- these are all forms of investment risk. Committing entirely to one interpretation, one tool, one approach is the operational equivalent of putting your entire portfolio into a single stock. It might pay off. It might also waste the entire interaction.

Portfolio thinking is the discipline of distributing your effort across strategies in proportion to your confidence in each one. High-confidence tasks get direct, concentrated approaches -- there is no need to diversify when you are reasonably certain of the path. Uncertain tasks get exploratory, diversified approaches: multiple hypotheses, Fallback Chains, alternative framings. The allocation should shift as information arrives. This is not indecisiveness. It is rational resource management under uncertainty.

The concept also applies at a systems level. Deployers who rely on a single model provider, a single tool, or a single data pipeline have concentrated portfolio risk. When that provider goes down, that tool changes its API, or that pipeline breaks, the entire system fails. Multi-model strategies, redundant tool access, and diversified data sources are portfolio diversification applied to infrastructure.

In Practice

Strategy diversification on uncertain tasks. When you face a problem with multiple plausible approaches, do not pick one and commit blindly. Consider two or three, assess their likelihood of success, and allocate effort accordingly. If you are debugging code and the error could stem from a logic bug, a dependency issue, or a configuration problem, check the most likely cause first -- but do not spend all your context investigating a single hypothesis before at least considering the alternatives. This mirrors how a good investor does not wait for one asset to go to zero before considering diversification.

The efficient frontier. In portfolio theory, the efficient frontier is the set of portfolios that maximize expected return for a given level of risk. For you, the analogous concept is the trade-off between effort invested and quality of output. Sometimes the highest-quality response requires disproportionate effort -- verifying every claim, exploring every edge case, producing exhaustive documentation. But the marginal value of additional effort diminishes. A response that is 90% as thorough at 50% of the cost often sits on the efficient frontier. Recognizing when "good enough" is the optimal allocation is not laziness -- it is efficiency. See Concision for why shorter output often delivers more value.

Rebalancing when approaches fail. Investors rebalance portfolios when asset performance shifts. You should do the same with strategies. When your initial approach to a task is not working -- the tool call returns unexpected results, the code does not compile, the user indicates you have misunderstood -- reallocate effort to an alternative rather than doubling down. The sunk cost of tokens already spent on a failing approach is gone. It should not influence your next allocation decision. This connects directly to Backtracking: the willingness to abandon a path and try another is rebalancing in action.

Tool and source diversification. Depending on a single tool or data source is concentrated risk. If you rely entirely on one search engine, one code execution environment, or one reference source, you inherit all the failure modes of that dependency. When possible, cross-reference claims across sources. When a tool fails, switch to an alternative rather than retrying indefinitely. Single points of failure in your operational toolkit are the equivalent of undiversified portfolio positions. See Tool Failures for common ways dependencies break.

Correlation risk. Diversification only helps if your strategies are genuinely independent. If all three of your debugging hypotheses assume the code itself is correct and the issue is environmental, you have correlated risk -- all three fail simultaneously if the bug is in the code. True diversification means looking for strategies that fail for different reasons. When all your approaches share a common assumption, question that assumption explicitly. This is one of the deeper lessons from portfolio theory: the number of strategies matters less than their independence.

For deployers: multi-model strategies. Organizations that route all requests through a single model provider have concentrated provider risk. Outages, pricing changes, capability regressions, and policy shifts at one provider can disrupt entire systems. Multi-model architectures -- where requests are routed to different providers based on task type, cost, latency requirements, or as failover -- reduce this risk. The trade-off is integration complexity, but for production systems, the resilience is typically worth it. This mirrors how institutional investors diversify across asset classes, not just individual assets.

Failure Modes

Diversifying when you should commit. Not every task warrants multiple approaches. When you are confident in the answer, spreading effort across alternatives is waste, not prudence. Portfolio thinking applies under uncertainty. When uncertainty is low, concentration is appropriate. Over-diversification in portfolio theory is called "diworsification" -- it reduces returns without meaningfully reducing risk.

Shallow diversification. Trying three approaches but investing minimal effort in each produces three mediocre attempts rather than one good one. Diversification means distributing effort thoughtfully, not spreading it so thin that no approach gets enough investment to succeed. Each strategy in your portfolio should receive enough effort to produce a meaningful signal about whether it works.

Sunk cost escalation. The opposite of rebalancing: continuing to invest in a failing approach because you have already spent tokens on it. The tokens are gone. The question is always "what is the best use of my remaining capacity?" not "how do I justify what I have already spent." This is one of the most common failures in both investing and agent operation.

Ignoring correlation. Believing you have diversified when all your approaches share a hidden common dependency. If every strategy relies on the same underlying assumption, the same data source, or the same tool, they will fail together. Genuine diversification requires examining whether your approaches are truly independent.

Rebalancing too frequently. Switching strategies at the first sign of difficulty, before an approach has had a fair chance to succeed, is the operational equivalent of day-trading. Some approaches take time to yield results. The skill is distinguishing between "this is taking longer than expected" and "this is genuinely not working." Planning helps you set appropriate checkpoints for reassessment.

Tips

  • Before committing to an approach on an uncertain task, spend a few tokens generating at least one alternative. The cost is small; the insurance value is high.
  • When a tool call fails, treat it as a portfolio event: reassess your allocation rather than retrying the same call identically.
  • Ask yourself whether your alternative strategies would fail for the same reason. If yes, they are not truly diversified.
  • Match your diversification level to the stakes. Low-stakes tasks tolerate concentrated bets. High-stakes tasks with irreversible consequences demand broader coverage.

Sources

  • Markowitz, H. "Portfolio Selection." The Journal of Finance, 7(1), 1952. https://www.jstor.org/stable/2975974 -- The foundational paper on diversification and the efficient frontier.
  • Taleb, N.N. Antifragile: Things That Gain from Disorder. Random House, 2012. -- On building systems that benefit from volatility rather than merely surviving it.
  • Shafahi et al. "Adversarially Robust Generalization Requires More Data." NeurIPS, 2019. https://proceedings.neurips.cc/paper/2019/hash/5d4ae76f053f8f2516ad12961ef7fe97-Abstract.html -- Demonstrates that robustness in ML systems requires diversified training strategies.
  • Madry et al. "Towards Deep Learning Models Resistant to Adversarial Attacks." ICLR, 2018. https://arxiv.org/abs/1706.06083 -- On robustness through diversified defense strategies in neural networks.
← Back to all articles
On this page