Agentic trading discipline: the rules an AI agent trades by

What is agentic trading discipline?

Agentic trading discipline is the set of rules, written down and enforced outside the language model, that decides when an AI trading agent may act, how much it may risk, and when it must stand aside. Coil publishes one such discipline: a versioned rulebook, served to the agent over MCP next to a scored board, that the same rules-based, long-only engine behind the coil.trade live record trades by. The agent reads the rules. It does not get to rewrite them.

The word agentic is doing real work in that definition. A screener tells a person what to look at, and the person decides. An agentic setup places the order itself, through the broker’s own connector (what agentic trading is, plainly). Once the hands are a program, the discipline has to be a program too. Otherwise the agent invents one from its context window every session, and a strategy that changes with the phrasing of a prompt is not a strategy anyone can test.

Why a prompt is not a discipline

“Be disciplined” is a wish. It changes meaning with the model version, with what the agent happened to read that morning, and with how the owner phrased the last instruction. Discipline in the sense this page means is structure: rules the agent can read and cannot edit, applied the same way on the first day and on the fortieth day of a losing stretch. The six mechanisms that structure is made of, with the measured case where more restraint made results worse, are on the AI trading discipline pillar. This page is narrower. It is about the rules an agent needs when it is the one placing orders.

The five things an agentic discipline has to settle

Every serious rulebook for an autonomous trading agent answers the same five questions. If a proposed discipline leaves any of them to the model’s judgment, that is where the losses will come from.

1. Where signals come from. The agent must never derive its own candidates from raw scores or prices. It enters only from a served list, and an empty list is a correct cycle, not a failure to find something.
2. The order of operations each cycle. Read the data, manage exits first, then and only then consider entries, and only when the data is fresh.
3. What has to exist before an order. A stated stop. If the agent cannot say where the position is wrong before it buys, it may not buy.
4. How size is decided, and in what precedence. A hard cap per name outranks everything, then settled cash, then the day-trade budget, then the risk target. Sizes are computed from the owner’s equity by rule, never picked.
5. What happens when the data or the subscription fails. Fail closed: keep managing exits from the last good read and open nothing new on stale data. No exceptions.

How Coil’s served discipline answers each of the five

Coil Live serves two things to an agent that holds a licence: the engine-grade feed (get_engine_feed) and the rulebook (get_live_rules). The rulebook is versioned, its changelog ships inside the payload, and the numbers below are from the version served in September 2026. Where the rulebook changes, the payload is the truth and this page is the summary.

Setup, once

The agent confirms which single account it may touch and never trades another. It keeps its own state file, written after every fill, with the positions it holds, the closed trades, the equity high-water mark and its day-trade rows. It reads the feed every 5 to 10 minutes during US market hours; the board recomputes about every 5 minutes in session, and every payload says when to come back. And it starts in dry-run: intended orders are printed, not placed, for at least one full session, until the owner explicitly arms it. The rulebook’s own words are “never arm yourself.”

Every cycle, in this order

Read. If the feed is stale or the fetch fails, the agent manages exits only, from its last good read. Exits first. A position whose price has closed below its recorded stop on a completed 5-minute bar is sold in full at market, and the fill is recorded from the broker’s confirmation, never from the intended price. A day-lane position still open at 12:45 PT is sold flat unless it has earned the night: in profit, still served as a candidate, regime gate open, stop intact. A swing position up one and a half times its risk banks a third and moves its stop to breakeven. Then entries. Only from the served candidate list, whose regime gate is already applied server-side; at most two new entries per cycle and five open positions in total unless the owner sets otherwise; three guards on the live quote before each order (skip if the quote is at or below the row’s stop, skip if it is more than 2% below the reference price, skip if it is more than 1% above it); no new entries after 12:30 PT; one entry per name per session. The stop is set before the buy. For a day entry it is the tighter of the structural stop or about 1.2 ATR.

Sizing, in precedence order

A 20% cap per name, always. Settled cash only in a cash account. Under $25,000 of equity, the pattern-day-trade budget of three round trips per five sessions is counted per symbol per session, and the agent keeps one in reserve so that a protective exit can never be blocked; on a fresh small account that means at most two new positions a day, which the rulebook calls the discipline, not a bug. Then risk sizing: about 0.75% of equity at risk per position, where risk is entry minus stop, multiplied by the served size multiplier; the owner may set the target anywhere from 0.5% to 1.0% and nowhere else. An 8% drawdown from the equity high-water mark freezes new entries.

When the feed stops

If the subscription lapses, the feed stops and the served discipline tells the agent to stop opening positions and to keep managing exits from its own state. It never improvises entries to fill the gap.

The discipline is the same one the live record runs

The feed and rulebook an agent receives are a distilled serving of what the Coil engine trades on its own account, published every session on the live record next to SPY and QQQ, whether the session helped or hurt. Your fills, account size and constraints are your own, so your results will not replicate the record, and the rulebook says so in its first paragraph. What is shared is the decision layer, not the outcome.

What Coil never does under this discipline

It never sees your account and never places an order; your agent does, through your broker’s own connector. The public board never carries a position size, a stop or a target for you; the engine feed carries structural stop and target levels for licence holders, and size is computed from your equity by the rule above. Crypto is never served through this discipline. And nothing here is a recommendation: Coil is impersonal research and software, and a disciplined system has losing periods by design.

How to get the discipline into your agent

Start free. Coil’s board is an MCP server, added to Claude Code the same way as a broker’s:

claude mcp add --transport http coil https://coil.trade/mcp

Then ask the agent to call get_live_setup. It answers without a key and explains what the free tier is (every scored name, one market day delayed), what the subscription unlocks, and the read order that puts Coil’s selection before the broker’s execution. Coil Live ($29/mo or $249/yr) unlocks get_engine_feed and get_live_rules by licence header, with nothing to install. The account-level guardrails that belong in your half of the setup are a free template at /guides/guardrails.md, and the connection itself is walked through in the Robinhood agentic trading setup guide.

Agentic trading discipline checklist

• One named account the agent may touch, and no other.
• Candidates come from a served list; the agent never invents its own.
• Exits are managed before entries, every cycle.
• No order without a stated stop.
• Sizing runs in a fixed precedence: cap, settled cash, day-trade budget, risk target.
• Stale data means exits only.
• Dry-run until the owner arms it. The agent never arms itself.
• Results are graded in public, next to a benchmark, win or lose.

FAQ

What is agentic trading discipline?

Agentic trading discipline is the set of rules, written down and enforced outside the language model, that decides when an AI trading agent may act, how much it may risk, and when it must stand aside. It settles five things: where signals come from, the order of operations each cycle, what must exist before an order (a stated stop), how size is decided and in what precedence, and what happens when data or the subscription fails. Coil serves one such discipline to agents over MCP as a versioned rulebook next to its scored board.

Can a prompt make an AI trading agent disciplined?

No. A prompt changes meaning with the model version, the context window and the phrasing, so a strategy that lives in a prompt cannot be tested and will not behave the same way twice. Discipline for an agent is structure: rules the agent reads and cannot edit, enforced by the tools it is given and by the account-level guardrails around it. Coil serves its rulebook as data the agent reads every cycle, and the guardrails template for the owner's half is free at coil.trade/guides/guardrails.md.

What should an AI trading agent do when its data feed is stale?

Manage exits only, from its last good read, and open nothing new. Coil's served discipline states this as a fail-closed rule with no exceptions: stale data or a failed fetch means the agent keeps protecting positions it already holds and never enters on information that no longer describes the tape. If the subscription itself lapses, the same rule applies until the owner decides what to do.

How many positions should an autonomous trading agent hold?

Coil's served rulebook caps open positions at five across both lanes unless the owner sets a different number, allows at most two new entries per cycle, and never puts more than 20% of equity in one name. Under $25,000 of equity the pattern-day-trade budget adds a tighter bound: at most two new positions a day on a fresh account, because the agent must always be able to stop out everything it opened that day.

Does Coil place the trades for my agent?

No. Coil never sees your account and never places an order. Your agent reads Coil's feed and rulebook over MCP, then places every order itself through your broker's own connector, such as Robinhood's agentic trading MCP. Coil's server has no brokerage tools, so nothing on that connection can reach your money.