The Executor

The algorithm is the soldier in the field; the language agent is the general at headquarters. That is the whole design in one sentence, and everything below is the consequence of taking it seriously: the soldier is fast and literal, the general is slow and contextual, and neither of them is allowed anywhere near the risk limits.

Layer 1 — the soldier

Execution sits entirely in code. It connects to the market feed, acts in milliseconds, and places every order with a take profit and a stop loss whose slippage buffer was computed from history for that instrument at that hour — not from a round number someone liked. It carries its own circuit breakers: an anomalous move closes to market and blocks new entries immediately, without waiting for anything upstream to have an opinion.

What it does not do is decide size. It receives that from the layer above. It does not read news, it does not know what a central bank is, and it does not improvise. Rules only.

Layer 2 — the guardian

Every order passes through the guardian before it exists. Position sizing comes from a fixed percentage of risk adjusted by realized volatility. Total exposure is aggregated across strategies, and a correlation matrix collapses instruments that are the same bet wearing different tickers — two US index contracts are one risk, not two.

Then the drawdown cascade, which is the part worth memorizing:

• A drawdown of daily drawdown limit: 2% in a day and no new positions open for the rest of it.
• A drawdown of weekly drawdown limit: 4% in a week and every size is halved.
• A drawdown of monthly drawdown limit: 8% in a month and everything pauses until a person reviews it.

The guardian holds a mathematical veto over size and over execution, and it does not depend on a language model for any of it. That is the point of putting it in code: a limit that can be talked out of is not a limit.

Layer 3 — the general

Every strategic review cadence, minutes: 15 minutes, the strategic layer gathers price structure, volume, order flow, news and sentiment, scores each source and deduplicates by event rather than by article — five outlets running the same wire story are one piece of evidence, not five. Text is a modifier of the quantitative signal, never the signal itself. When narrative and hard data disagree, the outcome is not a coin flip: the position is flagged as conflicted and exposure is reduced.

Its adjustments do not go to execution. They go to the guardian first, which can shrink them or refuse them outright.

Why that cadence

A review every review cadence: 15 minutes instead of one per second. Language models take seconds to answer and must never sit in the execution path; a cadence measured in minutes keeps the cost of thinking proportionate to the value of thinking, and it is fast enough to catch a regime change and reconfigure tactical parameters, which is the only thing this layer is actually for.

Versioned events, not a config file

The channel between layers is never a flat configuration file. A shared file invites race conditions and has no history. Instead, every change is a versioned event — a timestamp and an incrementing number — in a lightweight transactional store. Execution consumes the latest valid event atomically, and the agent's reasoning is appended to an immutable log alongside it.

That log is a product requirement, not an operational nicety. It is what makes a post-mortem possible, and it is what the audit trail is made of.

The kill switch

Before real capital, the agent has to earn its place: a comparative backtest of the algorithm alone against the algorithm plus the agent, measured on Sharpe. If the second does not beat the first, the agent adds no value however well it argues. Then a paper period long enough to cover a quiet regime, a volatile one and a scheduled macro event. Then auditable logs.

And then the switch itself: if measured value turns negative, the system drops automatically into what we call brainless mode — Layer 1 and Layer 2, execution and risk, with the general switched off. Nobody has to notice first.

What is not guaranteed here

A stop loss is not a promise about price. Gaps, thin liquidity, flash crashes and slippage are real and the historical buffer only narrows them. The quality of the strategic layer is bounded by the quality of its sources, and no scoring scheme replaces measuring the whole agent empirically.