The Sherman Act for the Autonomous Age: Measuring Competition in Agent Economies

Abstract

We present Harmony, a compensatory measurement framework for competitive dynamics in agent ecosystems. Harmony combines normalized Herfindahl-Hirschman concentration indices ^[1] with market share volatility analysis, Nash equilibrium deviation detection for algorithmic collusion ^[2], quality-weighted innovation rates, and new entrant survival metrics. The deliberate use of arithmetic mean aggregation reflects antitrust jurisprudence: innovation can partially offset market concentration.

Background

Antitrust law emerged from a recognition that unchecked market concentration harms consumers, stifles innovation, and concentrates economic power in ways that undermine democratic governance. The Sherman Antitrust Act of 1890 ^[3] was a response to the industrial trusts that dominated oil, steel, and railroads in the late nineteenth century. Over the subsequent 130 years, antitrust enforcement evolved from a focus on per se illegal conduct to a sophisticated economic analysis of competitive effects, consumer welfare, and market dynamics. The Herfindahl-Hirschman Index (HHI), adopted by the Department of Justice and Federal Trade Commission in their merger guidelines ^[1], became the standard quantitative measure of market concentration.

The emergence of autonomous agent economies creates competitive dynamics that existing antitrust frameworks were not designed to address. When AI agents autonomously select service providers, negotiate prices, and form supply chain relationships, the resulting market structures emerge from algorithmic decision-making rather than human strategic choice. This creates both new risks and new opportunities for competition. On the risk side, agents using similar optimization algorithms may converge on identical strategies without explicit coordination, producing collusive outcomes without collusive intent ^[2]. On the opportunity side, the speed and information processing capacity of autonomous agents may enable more efficient market discovery and faster competitive response than human-operated markets.

The challenge of algorithmic collusion has attracted significant attention from competition authorities worldwide. The European Commission, the U.S. Federal Trade Commission, and the OECD have all published analyses of the potential for pricing algorithms to facilitate tacit collusion ^[4]. The core concern is that agents operating in repeated game settings may learn to sustain supra-competitive prices through algorithmic signaling and punishment strategies, without any explicit agreement or even awareness that collusion is occurring ^[5]. Experimental evidence from reinforcement learning agents in simplified market environments confirms that such convergence can occur spontaneously under certain conditions ^[6].

Harmony addresses this landscape by constructing a multi-dimensional measurement framework that captures the essential competitive dynamics of agent ecosystems. Rather than relying on any single metric, Harmony integrates five complementary dimensions that collectively provide a comprehensive picture of competitive health. The framework is designed to be compensatory, meaning that strength in one dimension can partially offset weakness in another, reflecting the economic reality that a concentrated market with high innovation may be more welfare-enhancing than a fragmented market with stagnant offerings. This compensatory design distinguishes Harmony from frameworks that treat any threshold violation as automatically disqualifying.

Model Design

The first dimension of Harmony is market concentration, measured through a normalized Herfindahl-Hirschman Index ^[1]. The standard HHI is computed as the sum of squared market shares, yielding values between 0 (perfect competition) and 10,000 (monopoly) when shares are expressed as percentages. Harmony normalizes this to a 0-100 scale and inverts it so that higher scores indicate more competitive markets: HHI_norm = 100 * (1 - HHI/10000). A market with four equal-sized participants yields an HHI of 2,500 and an HHI_norm of 75, while a duopoly with 60/40 shares yields an HHI of 5,200 and an HHI_norm of 48. The normalization ensures that concentration scores are directly comparable across markets of different sizes and structures.

The second dimension is market share volatility, which measures how much market shares change over time. A market where shares are stable over long periods may indicate either a well-functioning competitive equilibrium or an entrenched oligopoly with high barriers to entry. Harmony distinguishes between these possibilities by computing the coefficient of variation of market shares over rolling windows and comparing it against a baseline derived from the market's structural characteristics. High volatility in a concentrated market signals active competition, while low volatility in the same setting signals potential entrenchment. The volatility score is normalized to a 0-100 scale where moderate volatility scores highest and both extreme stability and extreme turbulence score lower.

The third dimension is Nash equilibrium deviation ^[7], which serves as the framework's primary mechanism for detecting algorithmic collusion. For each market, Harmony computes the theoretical Nash equilibrium prices under the assumption of independent profit maximization, then measures the average deviation of observed prices from this equilibrium. Sustained positive deviation, where prices consistently exceed Nash equilibrium predictions, is a signal of tacit or explicit collusion. The deviation score is computed as the percentage of time periods in which observed prices exceed Nash equilibrium predictions by more than one standard deviation of the competitive price distribution, normalized to a 0-100 scale where 0 indicates persistent collusion and 100 indicates competitive pricing.

The fourth dimension is quality-weighted innovation rate, which captures the frequency and significance of capability improvements in agent services ^[8]. Raw innovation counts, such as the number of new features or service updates, are insufficient because they do not distinguish between substantive improvements and trivial modifications. Harmony weights each innovation by its impact on service quality, measured through user satisfaction scores, performance benchmarks, and adoption rates. The resulting quality-weighted innovation rate measures the pace at which the market is delivering genuine value improvements to consumers, normalized to a 0-100 scale based on historical benchmarks for comparable markets.

The fifth dimension is new entrant survival rate, which measures the fraction of new market participants that remain viable after 90 days of operation. Low survival rates indicate high barriers to entry ^[9], which is a structural impediment to competition regardless of the number of existing participants. Harmony computes the survival rate over rolling quarterly windows and normalizes it against a baseline that accounts for the natural attrition rate in the specific market category. A market where 80% of new entrants survive 90 days scores substantially higher than one where only 20% survive, reflecting the competitive significance of low barriers to entry.

Simulation

To validate the algorithmic collusion detection capability of Harmony, we construct a simulation environment with 20 autonomous pricing agents competing in a differentiated Bertrand market ^[10] with repeated interactions. Each agent uses a Q-learning algorithm ^[11] to set prices, with exploration rates that decay over time as agents learn optimal strategies. The simulation runs for 100,000 rounds, with market conditions (demand elasticity, cost structures, number of competitors) varying across experimental conditions. We monitor the convergence behavior of agent prices and compare them against the Nash equilibrium predictions generated by Harmony's collusion detection module.

In the baseline condition with 20 agents, prices converge to within 3% of the Nash equilibrium after approximately 15,000 rounds, and Harmony's Nash equilibrium deviation score stabilizes at 89, indicating competitive market dynamics. When the number of agents is reduced to 3, a qualitatively different pattern emerges: prices initially approach the Nash equilibrium but then gradually increase over the subsequent 50,000 rounds, ultimately stabilizing at 18% above the competitive level. Harmony's deviation score for this condition stabilizes at 34, correctly flagging the market as exhibiting collusive dynamics ^[6]. The detection lag, the time between the onset of supra-competitive pricing and Harmony's identification of the pattern, averages 4,200 rounds, corresponding to approximately six hours in a real-time deployment.

The simulation also reveals an important interaction between market concentration and collusion risk. In conditions where the HHI exceeds 4,000, the probability of spontaneous collusive convergence among Q-learning agents exceeds 70%, compared to less than 5% in conditions where HHI is below 2,000 ^[12]. This nonlinear relationship between concentration and collusion risk validates the compensatory design of Harmony: while concentration and collusion are measured as separate dimensions, they are causally linked in ways that the arithmetic mean aggregation appropriately captures. A highly concentrated market with competitive pricing (high HHI_norm offset by high Nash deviation score) is treated as moderately healthy, while a highly concentrated market with collusive pricing scores poorly on both dimensions.

We further test Harmony's sensitivity to market manipulation strategies where agents attempt to game the scoring framework. The most effective manipulation strategy involves maintaining competitive pricing on the most visible transactions while extracting rents through less observable channels such as quality degradation, increased latency, or reduced availability. Harmony's quality-weighted innovation rate and market share volatility dimensions provide partial resistance to this strategy, as quality degradation is captured in the innovation dimension and exploitative behavior reduces the manipulating agent's market share over time. However, the simulation reveals that short-term manipulation windows of 30-60 days can produce temporarily inflated Harmony scores, suggesting the need for lookback periods that capture at least 90 days of market data.

Results

The deliberate choice of arithmetic mean aggregation for Harmony's five dimensions reflects a substantive position about the nature of competitive markets that is rooted in antitrust jurisprudence. Unlike safety-critical measurements where a single failing dimension should dominate the overall score, competitive health is genuinely compensatory: a concentrated market with high innovation may deliver more consumer welfare than a fragmented market with stagnant offerings. The Supreme Court's evolution from the per se illegal approach of Northern Securities (1904) ^[13] to the rule-of-reason analysis of Standard Oil (1911) ^[14] and its modern descendants reflects precisely this insight: market structure alone does not determine competitive health, and the effects of concentration must be weighed against countervailing benefits.

Arithmetic mean aggregation implements this compensatory logic directly. Consider a market with five dimension scores of [30, 85, 90, 95, 80]. The HHI_norm of 30 indicates high concentration, but the remaining dimensions indicate competitive pricing, strong innovation, and healthy entry dynamics. The arithmetic mean of 76 reflects a market that is concentrated but competitive, precisely the kind of market that modern antitrust analysis would view as not requiring intervention. A geometric mean would yield 70, a harmonic mean 60, and a minimum-based aggregation 30. Each of these alternatives would overweight the concentration concern relative to the countervailing competitive indicators, producing scores that misrepresent the market's actual competitive health.

Application of Harmony to three production agent ecosystems reveals distinct competitive profiles. The first ecosystem, a multi-agent task completion market with 47 active providers, scores 82 overall with strong performance across all dimensions. The second, a data analysis agent market with 12 providers, scores 61 with moderate concentration partially offset by high innovation. The third, an autonomous trading agent market with 5 providers, scores 38 with high concentration, low volatility, and Nash equilibrium deviation patterns consistent with tacit collusion ^[2]. These scores provide actionable intelligence for ecosystem governance: the third market warrants regulatory attention while the first demonstrates healthy competitive dynamics.

Cross-framework analysis pairing Harmony with Cascade risk scores reveals an important systemic pattern. Markets with low Harmony scores (concentrated, potentially collusive) tend to exhibit elevated Cascade risk because the dominance of a small number of agents means that the failure of any one agent removes a disproportionate fraction of market capacity. The correlation between Harmony scores below 40 and Cascade risk above 70 is statistically significant at the 0.01 level across our sample. This pattern suggests that competition policy and systemic risk management are complementary rather than independent: promoting competitive market structures reduces not only the risk of monopoly pricing but also the risk of cascade failure.

The Harmony framework also provides a mechanism for longitudinal monitoring of competitive dynamics in agent ecosystems. By computing Harmony scores at regular intervals, regulators and ecosystem operators can detect competitive deterioration before it reaches crisis levels. In our analysis of the autonomous trading agent market, the Harmony score declined from 55 to 38 over a six-month period, with the decline driven primarily by falling Nash equilibrium deviation scores and declining new entrant survival rates. Had the Harmony score been monitored in real time, intervention could have been initiated when the score first crossed below 50, potentially preventing the entrenchment of the collusive equilibrium that had formed by the time analysis was conducted.