A software engineer has built an AI trading agent that proposes changes to its own rulebook, but the agent cannot simply rewrite its instructions: each proposed rule must first pass a test on 20 unseen past trading days. The approach offers a concrete answer to a familiar problem with AI agents: how to learn from mistakes without turning every bad outcome into a new rule.
Watch Desk analysis
What happened
AWS software engineer Ji “Franco” Lu’s independently developed agent reviews its trades nightly. It checks whether the agent followed its plan, traded too often or held a position through an event it should have avoided before considering whether a trade made money. A loss following the rules is treated as bad luck, not automatic grounds for changing them.
When the review identifies a pattern, the agent can propose a rule. A copy of the agent then replays 20 past trading days it has not seen, and the rule must improve a selected measure without worsening others, such as the depth of its worst run of losses. Lu reviews proposed rules himself and accepts no more than two a week. Accepted rules are retested every five sessions and can be retired. Read the Digital Journal account.
Why it matters
This is a practical example of putting tests and human approval between an AI agent’s own suggestions and changes to how it behaves. The method also keeps rejected proposals, with reasons, so the agent can consult them rather than repeatedly revisiting failed ideas.
In one example described in the article, the agent proposed waiting 45 minutes after the market opened. On the 20 replayed days, the rule removed 11 losing early trades and gave up two profitable ones; its worst run of losses shrank by about a third. Another proposed rule reduced small losses but roughly doubled the worst losing streak, so it was rejected. These are results from the reported historical tests, not proof of future trading performance.
Our read
The interesting idea is not that an agent has discovered a way to beat the market. It is that a suggested change can be tested, tracked and rejected before it becomes part of the live rulebook. That is a much more useful story about agent autonomy than the usual promise that everything will be fine because the agent has a very confident tone.
Lu says replaying past days cannot guarantee how live trading will go, and the test days lose value if reused too often. His approach is to keep testing accepted rules and replace old test days with new ones. For anyone building agents, the transferable lesson is to keep a record of why rules exist, test changes against held-out examples and leave approval with a person.
What to watch
- Whether the rulebook continues to shrink or change as old rules are retested.
- How the approach performs when tested on new trading days rather than reused examples.
- Whether similar test-and-review methods prove useful for agents doing work beyond trading.
Discussion spark: Should an AI agent be allowed to propose and test changes to its own rules, or should even the proposal stage remain under human direction?
Sources and evidence
- An AI agent writes its own rules. Each one has to prove itself first. – Digital Journal (20 November 2026, 16:21 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.