When an AI trades and competes in a simulated market, what makes it truly learn and change itself?

This explores what it takes for an LLM agent acting in a simulated economy to count as self-evolving, meaning it changes its own behavior over time instead of executing a fixed strategy. The collection has no note on economic simulations specifically, so what follows applies its general findings on self-improving agents to that setting. The economic-simulation part is my inference, not something the notes state.

First, what actually changes. Self-improving agents update in two ways: a slow loop that rewrites the model's weights, and a fast loop that rewrites prompts, memory and tools (Do self-improving agents really split into two distinct loops?). Most recent progress is in the fast loop because scaffold edits are cheap and reversible. An agent in a market simulation is self-evolving in this sense when it revises its own playbook between rounds. It might rewrite its strategy notes, add a pricing tool, or record which counterparties burned it. More capable models don't automatically do this better. The ability to write useful edits stays flat across model tiers, but the ability to benefit from them peaks in mid-tier models (Do stronger models always evolve harnesses better?).

Second, what pushes the change. Self-improvement in isolation stalls. It runs into the gap between generating and verifying, loses diversity, and invites reward hacking. Reliable methods work by bringing in an outside anchor such as past model versions, third-party judges, user corrections or tool feedback (Can models reliably improve themselves without external feedback?). Profit, losses and rivals' responses could plausibly play that anchoring role in an economy. The collection also says single-agent improvement stalls in static settings, while co-evolution supplies pressure by adding dynamic peers, then adaptive environments and feedback, and finally an evolving mechanism for evolution itself (Can agents evolve beyond the constraints humans engineer?). A market full of other adapting agents is that kind of setting, and it may be why an economy suits this better than a fixed benchmark does.

The furthest-reaching version lets the agent change what it is optimizing for. One system has an outer loop propose new objectives and compile them into scoring code the inner loop uses immediately (Can agents evolve their own objectives during search?). Another co-evolves the evaluator alongside the agent, because fixed criteria saturate and get gamed as agents improve (Can evaluators improve alongside the agents they score?, Why do fixed benchmarks fail as agents grow stronger?). An agent that chases a fixed profit number will find exploits. One whose success criteria move is harder to game. Two practical requirements follow. The agent has to keep running benchmark, edit and re-test cycles, since persistence predicted long-horizon success better than initial quality (What predicts success in ultra-long-horizon agent tasks?). And its edits should be versioned and reversible, so you can tell which change helped and roll back the ones that didn't (How can agent self-evolution be made safe and auditable?).

Two caveats in the collection cut against taking simulated self-evolution at face value. Agents that beat abstract contests still fail on real occupational workflows, so success in a sandbox economy may not predict economic usefulness (Why do agent benchmarks not predict real economic value?). And simulated agents that change after life events do so generically, more uniformly and less person-specifically than humans (Do personality-conditioned agents change like humans do?). An agent that visibly changes in a simulation may be drifting in a generic way, not adapting in a way that matches a real economic actor. On the positive side, automated evolution can reach parity with human-designed agents on held-out tasks (Does automated evolution match human-built agent performance?), so the mechanism itself works.