Blockchain Systemic Risk: When Autonomous Agents Outrun the System
On April 2, 2026, the IMF published a note signed by Tobias Adrian that comes as a wake-up call: the tokenization of financial assets is not a marginal improvement of infrastructure. It is a structural reconfiguration of global finance, and it carries systemic vulnerabilities that regulators do not yet master.
But the IMF stops where the question becomes truly interesting.
I) What the IMF saw
The note identifies four distinct risks. The one that stands out the most, when taking a systemic view of blockchain, is the first: speed as a crisis amplifier.
In the traditional financial system, the two-day settlement delay (T+2) is not an inefficiency to be corrected. It is a discretionary buffer.
It gives central banks, clearing houses, and human actors time to detect a problem before it spreads. Tokenization has removed it. Transactions settle in an atomic, instantaneous, final, irreversible manner. The IMF writes that crises “will likely unfold faster, leaving less time for discretionary intervention.”
This is correct, at least today.
II ) What the IMF implicitly raises
The IMF diagnoses speed risk at the macroeconomic level. What it does not say is that this risk operates continuously, at the level of each transaction, long before regulators can intervene. This is already a problem today when humans execute transactions; it is amplified with the deployment of deterministic (programmed) agents, and it will intensify further with the deployment of AI agents on blockchain networks.
An autonomous agent executing an arbitrage strategy on a layer 2 blockchain (rollup) at 3 a.m. does not wait for a central bank announcement. It acts. And if it acts without perception, it acts blindly.
Developers account for almost everything, and AI agents are structured, architected, and equipped with bridges that allow them to fetch near-instant information to secure their transactions, based on their own algorithms.
But is this information reliable? The global system and its interdependencies evolve according to the actions of the agents themselves.
Their main challenge is to position themselves relative to the nominal regime of this environment. This regime is shifting in the short, medium, and long term.
This is where the IMF’s four risks converge into a single operational problem: the absence of a definition of the nominal.
III) Nominal: the concept no one defines
In materials science, one does not describe a bridge as “stable” in absolute terms. One defines its nominal behavior range, the conditions under which it deforms in a predictable and reversible way. Outside this range, the material fails.
Blockchain is no different. Ethereum, Solana, Avalanche, layer 2 blockchains, cross-chain bridges, oracles, etc., all have a nominal behavior: a state in which gas fees, finality times, revert rates, and bridge delays fluctuate within predictable bounds. This nominal is not a fixed value. It is a dynamic distribution, continuously measured over sliding windows.
When the IMF refers to “fragmentation of liquidity into digital silos” or “cascading liquidations driven by smart contracts,” it is describing events of departure from the nominal, a divergence. Moments when the infrastructure behaves differently from what agents, human or autonomous, had anticipated.
IV) The agentic tragedy
Here is the scenario the IMF hints at without explicitly naming it.
In 2026–2027, thousands of autonomous agents execute strategies across tokenized networks. Each is individually rational. Each optimizes for itself, based on its algorithmic architecture. But their synchronized behavior, thousands of agents detecting the same arbitrage opportunity at the same microsecond, is itself the cause of the stress they seek to identify or predict.
Individual intelligence produces collective degradation because it is blind to the nominal and to system invariants.
The response to this problem is partly regulatory, but not only. It is infrastructural: each agent must have access to a shared measurement of the network state, a signal that reflects the aggregated pressure of agentic activity on the infrastructure.
The blockchain infrastructure deforms under agentic load; it adapts, modifies its own rules to adapt, and so on.
Like a road connecting two medium-sized cities (two blockchains).
This road is built for cities of a few thousand inhabitants.
The road, initially designed to connect two cities, changes their rules, their “operating system.” It facilitates exchanges and increases their attractiveness. In any case, both cities eventually have more inhabitants (autonomous agents), and, more or less, the road that originally had a nominal regime of a few cars ends up being congested 90% of the time, its new nominal. These traffic jams become so well known to users, this nominal being known, that some begin to take alternative routes, which in turn are developed, which in turn further drive the development of the two cities, and so on.
I invite the reader to explore Thinking in Systems: A Primer by Donella H. Meadows, extraordinary when tackling the understanding of complex systems. She also quotes Paul Anderson: “I have never seen a problem, however complicated, that when approached correctly does not become even more complicated.”
I believe this perspective describes extraordinarily well the blockchain environment and its future with autonomous agents.
When the IMF says that “crises will unfold faster,” it describes a world where the nominal must be known before action, not after the crisis. When it refers to “liquidations driven by smart contracts without possible human intervention,” it describes a world where agents need an exogenous signal of the network state, because no human will be there to apply the brakes.
This signal, this nominal, must be constructed and cross-validated.
Like decentralization, where computation is distributed across multiple nodes and the results are compared.
The nominal could be defined in this way.
But it is still not that simple: there is a key problem raised at the very beginning of this reflection, time. Banks prefer to take time for analysis. Autonomous agents have a strong preference for immediacy.
In addition to a shifting nominal, one must define short-term, medium-term, and long-term nominals, and above all extract divergence at high speed.
And the reasoning continues: when do we consider that this divergence is noise or a truly critical event?
This introduces the notion of calibration. Thank you… Paul Anderson 🙂
V) The paradox of stability
One final observation, which came to me while reading the IMF note.
One might think that if L1/L2 networks become more performant, higher throughput, reduced latency, consensus bugs eliminated, the need to measure network state decreases. The opposite is true.
The more performant the networks, the more active the agents. The more active the agents, the more frequent, brief, and difficult to detect stress events become without continuous measurement. The average stability of a highly active network hides very high instantaneous variance.
A six-lane highway can be in perfect condition and completely congested at 8 a.m. on a Monday. The quality of the road tells you nothing about traffic.
I also introduce here the notion of context and narrative in blockchain.
I like to describe narrative as the metrics related to price, gas, etc.
But context? Context is the infrastructure.
You are a developer, and using Claude and LangGraph, you create a state-of-the-art AI agent, a true Formula 1.
Well, context is the state of the infrastructure, the condition of the road, in other words. If it is full of potholes, your Formula 1, even with a V12 engine, will not go very far.
VI ) What comes next
The IMF note is a strong institutional signal. It generally precedes coordinated regulatory requirements. Within the next twelve to twenty-four months, regulators will seek to define standards for tokenized asset infrastructure.
These standards will have to include an answer to a question that no one is yet clearly asking: how does an autonomous agent know the state of the infrastructure on which it operates?
Defining the nominal is not optional. It is the foundation of every secure autonomous decision.
