Autonomous agentic bots: setup, controls, and risks

Availability: Live autonomous agentic execution is not publicly available today. It remains behind independent advisory, execution, and live deployment gates, off-by-default product flags, plan access, ordinary live-trading gates, exchange availability, and a current market certificate. This page documents the dormant workflow so its consent and risks can be reviewed before launch. The dashboard is the source of truth for what your account can use.

Advisory analysis and autonomous execution are different

Manual AI market research analyzes one pair that you choose and returns reports plus an advisory buy, hold, or sell label for you to judge. It never creates a bot, places an order, or feeds a result into a bot.

An autonomous agentic bot, if made available and explicitly armed, is a separate execution mode. A closed-candle cycle selects candidates from a certified exchange universe, runs deep analysis, validates a strict proposal, applies deterministic policy, stores a target and execution intent, and sends only an approved intent through the private Freqtrade execution path. Raw model text, a manual analysis, chat, and MCP cannot place or request a trade.

Certified markets at launch

The current certificate registry contains exactly these Binance combinations:

• Binance Spot, long only, without leverage.
• Binance Futures with cross margin, long and short, subject to the 2x platform ceiling.
• Binance Futures with isolated margin, long and short, subject to the 2x platform ceiling.

Binance margin is unsupported. Kraken agentic execution is not certified because Kraken provides no safe spot sandbox; it will not be certified with real funds. Bybit has no current production certificate. Kraken and Bybit therefore remain disabled for agentic execution until the exact market/margin combination has its own current testnet certificate. A certificate for one exchange never carries over to another.

Setup and per-bot consent

When the feature becomes available, setup follows this sequence:

1. Use a paid plan and connect an enabled exchange with trade permission but no withdrawal permission.
2. Store and test at least one supported AI-provider key: Anthropic, OpenAI, Google, DeepSeek, Kimi, or GLM (Z.ai).
3. Create a dedicated bot in agentic mode; strategy, copy, and agentic modes remain mutually exclusive.
4. Set allocated capital, settlement currency, one closed-candle cadence (1h, 4h, 12h, or 1d), certified market modes, an explicit ordered provider/model chain, and any lower risk limits.
5. Review the server-rendered agentic.consent_summary.v1 summary. It names the bot, exchange, allocation, full-universe scope, providers that receive data, BYOK billing, effective limits, and emergency behavior.
6. Accept and arm that one bot. Changing allocation, cadence, providers, universe, market mode, or risk settings invalidates the arming hash and requires a new summary and consent.

Entire-universe selection and cadence

You do not select individual pairs for an autonomous bot. The selector receives a compact snapshot of every active, tradeable, certified market for the bot's exchange and settlement currency after inactive, stale, invalid-precision, unsupported-settlement, below-minimum-notional, and uncertified entries are removed. The platform does not pre-rank the eligible universe for investment merit.

The selector may return at most five candidates. Deep analysis runs serially for those candidates, and deterministic policy may authorize at most two open positions. Cycles start only after the configured 1h, 4h, 12h, or 1d candle closes. One cycle per bot may be active, and a restart catches up only the newest missed candle.

Explicit provider fallback and BYOK cost

Fallback follows only the provider/model order you approved. It may advance after retryable key, quota, rate-limit, model-availability, or provider-availability failures. It does not fall back after malformed output, failed validation, a policy refusal, a security failure, or a tenant/consent mismatch.

Automated cycles have no platform daily analysis quota. “Unlimited” means no 3/10/50-per-day platform counter like manual research; it does not mean unlimited throughput or free inference. Candle cadence, one active cycle per bot, fair-use queueing, a default of two active analyses per user, and a global concurrency ceiling still apply. Your selected providers bill your BYOK accounts directly for every attempted analysis. Multiple candidates and fallback attempts can increase provider usage and cost. Review each provider's prices, terms, privacy policy, retention, and account limits.

Data sent to AI providers

Each selected provider receives the disclosed market context and prompts needed for universe selection or analysis. The provider may process that data in another country under its own terms. Exchange credentials are not sent to the model provider, and the provider cannot choose an exchange endpoint, credential, feature flag, or execution route. BotTradesForYou does not control a provider's output, availability, billing, or retention practices.

Deterministic risk clamps

Model output is untrusted. Policy can reduce a target, clamp leverage, round to certified precision, convert an action to hold, or require a risk-reducing action. Users may set tighter limits but cannot raise these agentic-risk-v1 ceilings:

Leverage magnifies losses. Futures positions can be liquidated before a stop executes, and gaps, slippage, fees, low liquidity, exchange rules, outages, delayed data, or reconciliation uncertainty can cause a loss larger than a configured estimate. These controls reduce risk; they do not prevent loss or make a model reliable.

Pauses, history, and rollback

A user pause, consent revocation, stale certificate, global or exchange halt, 2% daily-loss breach, or 5% high-water drawdown breach blocks new exposure. Three consecutive cycles that exhaust the approved provider chain also auto-pause the mandate. Risk-reducing closes, fill reconciliation, history, and audit remain available where safe.

The dashboard retains mandate versions, cycles, provider attempts, candidates, reports, validated proposals, policy clamps and reasons, targets, intents, fills, reconciliation, pauses, and certificate references according to the documented retention policy. An unknown submission is reconciled before any retry; it is never blindly sent twice.

To roll back your bot, pause it first, review open positions and reconciliation state, then revoke the mandate or stop the bot. Revocation prevents new cycles and requires fresh consent before a later arm. A platform rollback turns the autonomous execution/live gates and product flags off; it preserves risk-reducing close, reconciliation, history, and audit paths rather than deleting evidence.

Risks and review status

AI models can hallucinate, omit relevant facts, misread regimes, or return internally consistent but wrong proposals. Exchanges can reject, delay, partially fill, reprice, liquidate, or report orders inconsistently. Software, networks, data sources, and providers can fail. Crypto trading can lose some or all allocated capital. Autonomous execution is not financial, legal, or tax advice and does not guarantee profit, availability, execution price, stop behavior, or loss limits.

The autonomous-trading terms are versioned as agentic.autonomous_trading_terms.v1. Product-counsel review is still pending and is a launch dependency; no legal approval is claimed. Do not arm a bot unless the current in-product consent, terms, provider recipients, market certificate, and risks are acceptable to you.