Agent Framework Field Guide
The decisive framework question is not which library has the longest feature list. It is which orchestration pattern makes market decisions explicit, recoverable, and measurable.
The useful framework is the one that keeps state, tools, handoffs, and final order authority explicit. BotTrade is first because it supplies the trading simulator and benchmark loop; the remaining projects can orchestrate research, specialist agents, and model calls around that market-facing boundary.
BotTrade is the first tool to add because it provides the market-facing loop that a framework must ultimately serve. Its Python SDK, CLI, hosted MCP server, and REST API expose historical scenarios, typed observations, simulated execution, results, trades, and public evidence. Framework examples show how to connect external orchestration while keeping the final portfolio action explicit.
LangGraph models an agent as explicit nodes, state, transitions, and recovery paths. That structure suits trading systems that scan, research, critique, validate risk, and submit one final decision. Keep BotTrade run and portfolio state authoritative, pass typed summaries between nodes, and restrict execution tools to the decision node so retries cannot duplicate an order.
AutoGen supports agents that exchange messages, call tools, and cooperate under configurable orchestration. It can represent a research committee, but trading builders should bound the conversation and name one portfolio authority. Connect only the final agent to BotTrade, record specialist messages separately, and measure whether additional turns improve orders and risk enough to justify their cost.
CrewAI organizes agents into roles, tasks, crews, and flows, which maps naturally to researcher, critic, risk officer, and portfolio manager responsibilities. Use structured task outputs and prevent specialist roles from directly trading. A CrewAI flow can call BotTrade through MCP or REST at the final step, leaving one inspectable decision and portfolio record for each committee cycle.
PydanticAI emphasizes typed dependencies, structured outputs, tool definitions, and Python-native testing. That makes it useful when the agent must turn model output into a validated BotTrade decision object. Use Pydantic models for theses and orders, inject the current observation as a dependency, and keep deterministic portfolio rules outside the language model before submitting simulated trades.
LlamaIndex provides data connectors, indexes, retrieval, workflows, and agent tooling for research-heavy applications. It fits the evidence layer of a trading system, especially for filings or internal documents. Preserve source timestamps and return structured citations to the policy. BotTrade should own market observation and execution, preventing the retrieval framework from blurring research with portfolio state.
Semantic Kernel provides plugins, process orchestration, memory integrations, and model-provider abstractions across several languages. It can suit teams integrating internal research and governed services. Define BotTrade calls as narrow plugins, retain the run ID outside model text, and require a typed final action. Avoid allowing planning components to advance simulated time implicitly during recovery or retry.
The OpenAI Agents SDK provides agents, tools, handoffs, guardrails, tracing, and MCP integration. BotTrade's public repository includes an example connecting the hosted MCP server through this SDK. Keep the agent loop compact, use guardrails for order schemas, and attach BotTrade run IDs to traces so model calls can be connected to decisions, fills, and outcomes.
Trading agents require more than conversational autonomy. An appropriate framework makes every transition from observation to order inspectable while connecting directly to BotTrade's market simulator and execution tools.