Six identified risks. That is the core output from Fidelity Digital Assets' latest research report on AI agents operating on public blockchains. The institutional research arm flags a structural contradiction at the heart of the AI-crypto thesis: autonomous agents may systematically prefer closed, centralized infrastructure over the public networks that token valuations depend on. For any quantitative trader building exposure to AI-agent token narratives, this is a risk-model input worth calibrating.
The Centralization Divergence
The report identifies a fundamental architectural tension. AI agents optimizing for execution efficiency, latency, and cost will gravitate toward systems that minimize friction. Closed, centralized platforms offer deterministic settlement, lower variable costs, and tighter API integration — all variables that reduce slippage and execution variance in automated strategies.
Public blockchains, by contrast, introduce probabilistic finality, gas fee volatility, and mempool contention. For an algorithmic agent evaluating infrastructure on a pure cost-function basis, the math does not favor decentralization. The implication: the very autonomy that makes AI agents compelling as a narrative does not inherently drive demand for native tokens on public networks.
This is not a theoretical concern. Any backtest of agent-driven transaction volume on public chains must account for the probability that agents migrate to cheaper, faster rails once those rails become available at scale.
Volume Without Value Accrual
The second critical risk vector: transaction volume ≠ token value. Fidelity's report warns that high automated payment throughput may not translate into native token price support. The mechanism is straightforward.
Automated agents generate high-frequency, low-value transactions. On public chains, this produces fee revenue for validators but not necessarily sustained buy pressure on the native asset. If agents batch, compress, or route transactions through Layer-2 solutions or alternative settlement layers, the base-layer token captures diminishing marginal value from each additional unit of economic activity.
For traders pricing AI-agent tokens on the assumption that "more agents = more transactions = higher token demand," the model requires a direct value-accrual link. Fidelity's analysis suggests that link is weaker than consensus pricing implies. Standard deviation on this thesis is wide.
Practical Calibration
Three checkpoints for any AI-crypto allocation:
- Infrastructure preference tracking. Monitor where new AI-agent deployments actually settle. If the ratio of centralized-to-decentralized agent deployments trends above 0.7 over consecutive quarters, the public-chain thesis degrades materially.
- Fee capture ratio. Measure native token fee revenue as a percentage of total agent-generated transaction volume. A declining ratio signals value leakage to alternative layers.
- Agent count vs. token velocity. Rising agent count with flat or declining on-chain token velocity indicates agents are transacting without holding — a bearish signal for long-duration token positions.
The Fidelity report does not invalidate the AI-agent category. It recalibrates the risk-adjusted expected return. For systematic traders, the actionable output is clear: model centralization migration as a base case, not a tail risk. Any position sizing that assumes public-chain lock-in by default is under-hedged.