“AI trading agent” has become a catch-all phrase. Product pages use it for grid bots, indicator stacks, chatbots with exchange APIs, reinforcement learners, and multi-service desks. Without a taxonomy, teams buy the wrong complexity, mis-size risk, and evaluate the wrong metrics.
This article maps the main types of AI trading agents used in practice—especially around crypto markets—so you can name what you are building, what you are buying, and what must sit behind a risk gate. For the base definition, see What Is an AI Trading Agent?; for the bot boundary, see AI Trading Agent vs Trading Bot.
Why taxonomy matters
Labels drive design choices. A market-making inventory controller and an LLM research agent may both “trade,” but they differ in latency budget, failure modes, and evaluation. Taxonomy is not academic hair-splitting—it is how you prevent an LLM on a millisecond path and a rules bot with withdraw keys.
If you cannot name the type, you cannot name the risk.
Three axes of classification
1. Decision graph
How much of the control flow is fixed at design time versus adapted at runtime?
- Fixed — scripted branches, grids, thresholds
- Scored — models emit scores; rules still decide
- Policy — learned mapping from state to action (e.g., RL)
- Planning / tools — agent selects tools and playbooks dynamically
- Multi-agent — specialized roles coordinate under a supervisor
2. Input modality
- Market microstructure only — trades, books, funding, inventory
- Structured alternatives — on-chain metrics, open interest, factor panels
- Unstructured context — news, governance, docs, social (often via LLMs)
3. Mandate
- Directional alpha / signal capture
- Market making and inventory skew
- Arbitrage / relative value
- Research + human-approved execution
- Portfolio rebalance / risk reduction
Real systems are hybrids. Taxonomy names the dominant pattern so architecture and risk stay honest.
Type catalog
1. Rule / scripted automation
Fixed logic: indicators, grids, DCA schedules, rebalance bands. Often sold as an “AI bot” with no adaptive decision graph.
- Fits: stable edges, simple ops, hot paths that must stay deterministic
- Main risk: wrong parameters, regime change, silent data bugs
- Note: valuable baseline—not “lesser,” just not agentic by default
2. ML-enhanced signal bot
Supervised (or similar) models score direction, volatility, or microstructure features; execution and sizing remain rule-bound.
- Fits: quant-style alpha research with controlled rollout
- Main risk: leakage, train/serve skew, concept drift
- Eval: classic backtest hygiene plus shadow scoring — evaluation guide
3. Reinforcement / policy agent
A policy maps sequential state to actions (including hold/cancel) under costs and constraints. The “intelligence” is optimized for cumulative reward, not a single label.
- Fits: inventory control, sequential execution, simulated environments with good fidelity
- Main risk: sim-to-live gap, reward hacking, unsafe exploration if ever live-trained
- Note: production use usually freezes the policy and surrounds it with hard limits
4. LLM / tool-using agent
Language models plan, extract events, call tools (data, risk preview, docs), and emit structured intents. Strong on warm/cold paths; weak as pure price oracles. Deep dive: LLMs for Crypto Market Analysis.
- Fits: narrative risk, research copilots, playbook selection, incident synthesis
- Main risk: hallucination, prompt injection, model drift, latency/cost
- Hard rule: schemas + validators + independent risk gate—never free-form orders
5. Multi-agent trading system
Specialized agents (research, risk, execution, monitoring) coordinate. Powerful when interfaces are strict; expensive when chatty and unbounded.
- Fits: desks separating cognition from control at scale
- Main risk: coordination tax, unclear ownership of intents, cascading tool failures
- Architecture: reference layers still apply—multi-agent is a topology, not a free pass on risk
6. Market-making / inventory agent
Mandate-first hybrid: quotes and inventory skew under adverse selection and fee regimes. May use ML or RL inside, but success is measured in inventory, spread capture, and survival—not story quality.
- Fits: continuous markets with measurable microstructure
- Main risk: toxic flow, venue outages, inventory blowups
- Latency: hot path stays deterministic even if models inform skew
7. On-chain / autonomous execution agent
Agents that sign or submit on-chain actions (swaps, LP, liquidations) under custody and gas constraints. Autonomy collides with irreversible transactions and MEV.
- Fits: narrow, heavily simulated strategies with spend limits
- Main risk: key compromise, failed txs, MEV, contract risk
- Control: least-privilege keys, simulation, human thresholds for size
Comparison table
| Type | Latency fit | Autonomy | Eval difficulty | Risk surface |
|---|---|---|---|---|
| Rule / scripted | Hot | Low | Lower | Params, data, regime |
| ML signal bot | Hot–warm | Low–med | Medium | Leakage, drift |
| RL / policy | Hot–warm | Medium | High | Sim gap, reward design |
| LLM / tools | Warm–cold | Med–high | High | Hallucination, injection |
| Multi-agent | Warm–cold | High | Very high | Coordination, ownership |
| Market making | Hot | Varies | High | Inventory, toxic flow |
| On-chain exec | Warm | Med–high | High | Keys, MEV, irreversibility |
How to choose a type
- Write the mandate in one sentence (alpha, MM, research+approve, etc.).
- List inputs you truly have in production—not in a notebook fantasy.
- Pick the simplest decision graph that can express the edge.
- Match latency tier: hot paths stay boring; planners stay warm/cold.
- Confirm you can evaluate and kill the system independently of the model.
- Only then consider LLM tools or multi-agent topology.
The build roadmap assumes you pick a type in Step 0 scope—not after wiring withdraw permissions.
Anti-patterns
- Universal agent — one system that “does everything” with no mandate
- LLM on the hot path — language latency and nondeterminism in quote engines
- Type laundering — calling any bot “AI” without adaptive control flow
- Risk mismatch — multi-agent autonomy with bot-grade monitoring
- Eval mismatch — scoring an LLM research agent only on PnL of a toy fill model
Type selection does not replace a risk gate, kill switch, or paper mode. Complexity is a liability until measurement proves otherwise.
Frequently asked questions
What are the main types of AI trading agents?
Common types include rule/scripted automation (often mislabeled as AI), ML-enhanced signal bots, reinforcement or policy agents, LLM/tool-using agents, multi-agent systems, market-making or inventory agents, and on-chain autonomous execution agents. Classify by decision graph, inputs, and mandate.
Is a grid bot an AI trading agent?
Usually no. A classic grid or DCA bot is scripted automation with a fixed decision graph. It can be valuable, but calling it an AI trading agent without adaptive policy, learning, or tool-using control flow is mostly marketing.
What type of AI trading agent should beginners build first?
Start with rule automation or an ML-enhanced signal bot behind a hard risk gate and paper mode. Add LLM tool use or multi-agent designs only after data quality, risk, and evaluation loops are stable.
How is an LLM trading agent different from an RL trading agent?
An RL or policy agent learns a mapping from market/state features to actions in a sequential decision problem. An LLM tool-using agent plans with language models, calls tools, and emits structured intents—strong for research and playbooks, weak as a pure price oracle and usually unfit for hot-path latency.
Related: What Is an AI Trading Agent? · Agent vs Bot · Architecture · How to Build