In short: every machine that trades is identified, has a person or firm that answers for it, works within caps on what it may commit, and can be stopped. When machines cascade, markets halt, and trades that were clear errors are undone.
The first test case
Much of today’s trading is already done by machines, and more is done by machines trading with other machines. They act in thousandths of a second. When many of them read the same signal the same way, prices can collapse and recover within minutes, and people whose orders were caught in the fall bear the loss.
Finance is where the framework’s rules for autonomous agents meet their first large test. If agents can be made answerable, capped, and stoppable here, where every action is already recorded, the same rules can work elsewhere.
Four rules
1. Every trading agent is identified. Markets know when an order comes from a machine, and whose machine it is. An agent trading under someone else’s identity, or none, is barred.
2. Every agent has an answerable party. A named person or firm answers for what each agent does, and pays for the harm it causes, without anyone having to prove intent (responsibility address).
3. Caps on what an agent may commit. Each agent works within limits on the size and pace of what it may trade without a person’s approval, set by its owner and enforced by the market it trades on. An agent that reaches its cap stops.
4. Halts and unwinding. When prices move faster than a set rate, trading pauses so people can look. Trades caused by a clear malfunction can be reversed within a set window, under published rules, so a machine’s error does not become someone else’s loss.
Testing before release
A trading agent that will trade at scale is tested first against conditions that could make it fail: sudden price moves, other agents behaving strangely, its own data going stale. The test results go to the market it will trade on. An agent that cannot show it will stop when it should is not released.
In the transition
Markets already use some of these tools: pauses on sharp moves, some registration of trading firms, rules against manipulation. What is missing is the agent itself as the unit: identification of each agent rather than each firm, caps on each agent, and an answerable party for each one. As machine advisers and personal agents begin to trade for ordinary people, not only for large firms, the gap widens. These rules can start now, on the markets that already record every order.
Who resists, who backs it
| Rule | Resists | Backs |
|---|---|---|
| Identification of each agent | firms whose edge depends on staying anonymous | market operators, regulators, ordinary investors |
| Caps and halts | the fastest traders | pension funds, long-term investors, anyone caught in a sudden collapse |
| Liability without proof of intent | firms deploying agents at scale | people harmed by agent errors |
What stays open
- Where to set the line between a malfunction that is reversed and a bad trade that stands.
- How caps apply to personal agents trading small sums for millions of people at once.
- How markets in different countries coordinate halts when agents trade across all of them.