A security team closes a vulnerability. Attackers study the patch, change their tools, and search the boundary it revealed. Defenders add detection. Attackers change how they move. Each local improvement is real, but the contest does not settle into safety. It produces another round.
When adaptive systems compete, one system’s improvement changes the environment of the other. The response then changes the first system’s environment in return. Recursive improvement becomes an interaction rather than a private loop, and progress can look like continual motion with no durable lead.
Defense changes offense. Security improvements alter the field on which the next attack is designed. Photo by FlyD on Unsplash
Running to remain in place
Evolutionary biologist Leigh Van Valen used the Red Queen image—running to remain in place—to describe dynamics in which lineages must keep adapting because the organisms around them are adapting too. The metaphor fits many human competitions. A campaign improves voter targeting; its opponent improves counter-messaging. A spam filter learns a pattern; spammers vary the message. A military develops armor; another develops a weapon that defeats it.
Performance is relative. A company can improve its product and still lose ground if a rival improves faster. A defender can block more attacks while the total threat rises. Each participant receives feedback partly generated by the other participant’s learning, so yesterday’s successful strategy becomes today’s visible weakness.
A improves, changing B’s world. B adapts, changing A’s world. Neither controls the loop they create together.
Competition can be productive. Rival firms discover cheaper methods. Athletic opponents draw better performances from one another. Security researchers harden systems by exposing weaknesses before criminals exploit them. Rules, norms, and shared interests can turn the pressure of competition toward useful ends.
But local rationality can also create a globally foolish race. If every political campaign uses more alarming messages because alarm briefly captures attention, public trust may fall even while each campaign improves its metrics. If high-frequency traders spend heavily to shave tiny fractions of a second from execution, no participant can safely stop, though society may receive little benefit from the accumulated expense.
Changing the game
The usual self-improvement question asks how an agent can become more capable. Competitive systems add another: can participants improve the rules of interaction, or will any restraint be exploited by the first defector? Arms control, technical standards, sporting rules, market regulation, and professional ethics are attempts to modify the larger loop.
Shared constraints are also strategies: A speed limit in a race is not merely less capability. If it is enforceable, it can redirect effort from dangerous escalation toward reliability, skill, or another good the participants still value.
Cooperation doesn’t remove adaptation. Participants learn the rule, discover its edges, and sometimes capture the institution enforcing it. A stable agreement needs observation, credible consequences, and enough mutual benefit that compliance survives short-term temptation. It also needs revision, because an old rule can freeze the advantages of yesterday’s winners.
AI agents will enter many of these loops as tools and participants. Faster detection can answer faster fraud; automated negotiation can answer automated pricing; generated persuasion can answer generated skepticism. The risk is not only that one system becomes extremely capable. It is that many capable systems create an escalation none of their owners intended.
Competition separates two kinds of progress that are easy to confuse. A player can become stronger while the game becomes more wasteful, dangerous, or dishonest. Shared constraints can redirect the contest, but only when participants can see violations and believe defection will cost more than restraint.
That leaves a question no contest can answer for itself. A better player is better only relative to the game and its objective. Before asking how quickly an improvement loop can turn, we have to ask whether its target deserves pursuit.