TL;DR

A single large language model wired directly to a trading API is a solved problem the wrong way. The right architecture is multi-agent: a Screener that holds the halal gate, a Signal agent that generates entry hypotheses, a Risk agent that vetoes oversized positions, and an Execution agent that places only what survives. Separation of concerns is not just cleaner code — it is the difference between a compliant halal system and one quiet bug away from putting a haram coin into a user's account. We built HalalCrypto multi-agent on purpose.

The single-model failure mode

If you let one model do everything — read prices, judge halal status, size the position, fire the order — every prompt collision becomes a risk event. The model that just summarised a Reuters article in working memory is the same model now deciding whether to allocate $400 to a token. Its halal gate becomes statistical, not procedural. Its risk cap becomes a soft suggestion. One in a thousand decisions, the seams leak.

For a halal product, one in a thousand is not acceptable. The whole premise of an AAOIFI-aligned framework, with Saudi Permanent Committee for Ifta and leading Saudi Islamic banks guidance, is that a coin is either on the permitted list or it is not. There is no probabilistic halal.

The multi-agent split we use

Four specialists, each with a narrow, auditable job:

The Screener runs the 4-gate halal filter (riba, gharar, maysir, haram-sector) every 24 hours. Its only output is a permitted-coin list. It does not see prices. It does not know about positions. It cannot be talked out of a verdict.
The Signal agent generates entry hypotheses on the permitted list only. It can read on-chain flow, momentum, volatility, and order-book depth. It produces ranked candidates. It cannot, by API design, place orders.
The Risk agent is the veto. It checks position-size caps, per-tier exposure limits, drawdown state, and concurrency rules. If a Signal candidate breaches any rule, Risk kills it silently. Risk has no upside incentive — its only job is to say no.
The Execution agent places spot-only orders on the user's supported exchange account via the read+spot-only API key. It refuses any order that did not pass through Screener + Signal + Risk in sequence. If the chain is broken, the order does not fire.

Why the separation matters operationally

Imagine a token that was on yesterday's permitted list but introduced a CeFi lending product overnight. Under a single-model system, the fact-check happens in the same context window as the trade decision — and if the lending news did not surface in the prompt, the trade fires. Under multi-agent, the Screener runs first, on its own schedule, with its own data sources, before Signal even sees the token. By the time Signal looks at the universe, the token is gone.

This is the structural reason a halal crypto bot needs to be multi-agent: the halal gate has to live upstream of the trade gate, on a different clock, with a different prompt.

Audit trail as a side-effect

Multi-agent has a quiet side-benefit: every decision leaves a paper trail. Each agent writes its inputs and outputs to a local audit log. When a user asks "why did you not buy XYZ this week?" we can answer with a specific trace: Screener removed it on day 3 because of a reported lending integration; Signal never saw it on day 4–7. That is the kind of legibility a single-model bot cannot produce.

For a Shariah-sensitive product this matters more than for a generic trading bot. Permissibility questions rarely have a single answer; they have a process. A bot that can show the process is a bot a scholar can trust.

Tiered intelligence, identical screen

The same multi-agent skeleton runs across all three tiers. What differs is which Signal model is plugged in:

Conservative ($49/mo) uses a momentum-and-trend Signal with conservative entry thresholds. It is the default recommendation for new users — same halal screen, slower cadence, smaller position sizes.
Moderate ($69/mo) adds an on-chain flow Signal and a volatility regime detector.
Multi-X ($99/mo) adds a pyramid-entry Signal with staged partial-exit controls. These are execution rules, not advertised outcomes.

The Screener is identical across tiers. The halal floor does not get cheaper at the lower tier or laxer at the higher tier. That is intentional.

Failure modes we have seen and fixed

A few real ones from our development cycle, recorded so the next builder does not repeat them:

Cache poisoning across agents. Screener output was cached in shared memory; a stale 48-hour cache nearly let a delisted token re-enter the universe. Fix: per-agent caches with hard TTLs, and a cold-start verification on every run.
Retry-loop riba. Execution retried failed orders four times in 60 seconds. On a thinly-traded permitted coin, the retries pushed the price up and the agent kept buying into its own slippage. Fix: linear back-off with a hard ceiling and a sanity check against the original Signal price.
Implicit leverage. A Binance margin balance left over from a user's prior account showed up as available capital. The agent treated it as cash. Fix: explicit margin-account check at API-key install; refuse keys that have margin enabled.

Every one of these is the kind of bug a single-model system cannot easily debug — because the responsible code is everywhere. Multi-agent gives you a layer to point at.

What this means for the user

Practically, the user sees three things: a published methodology, a position log they can verify on their supported exchange, and the ability to challenge a coin's classification. They never see the four agents directly, and that is correct — the architecture is plumbing. Its job is to make the halal floor procedural rather than probabilistic.

Funds remain on the user's supported exchange account at all times. Card payments run through DodoPayments; crypto payments through NowPayments. We do not custody, we do not lend, we do not hypothecate.

Concluding paragraph

A multi-agent split is not the trendy answer; it is the boring, correct one. The halal screen has to live upstream of the trade. Risk has to be allowed to say no without arguing. Execution has to refuse to act on broken chains. None of that is glamorous, and none of it shows up in marketing copy — but it is the difference between a halal product that holds up under scrutiny and one that quietly does not. If you want to test the architecture for yourself, start at the Conservative tier.