The Axiom

Agents

18 pages

Agent Memory

Covers the four agent memory types (in-context, episodic, semantic, procedural) and how to implement each — getting this right separates agents that learn across sessions from ones that forget everything.

memoryagentsepisodicsemantic

Agent Memory Systems

Production memory infrastructure for AI agents — Mem0 (ecosystem breadth), Zep/Graphiti (temporal facts), and Letta (long-horizon paging) — with the dual-layer hot/cold architecture as the emerging production standard.

memorymem0zepletta

AutoGen / AG2

AG2 (formerly AutoGen) is Microsoft's event-driven multi-agent framework. GroupChat is its primary pattern — multiple agents in a shared conversation thread. Renamed from AutoGen in late 2024. Best for enterprise-scale, cross-language, and complex group coordination.

autogenag2multi-agentgroupchat

Computer Use

Computer Use lets Claude control a computer by observing screenshots and issuing mouse/keyboard actions in a tight loop — use only when structured APIs are unavailable, always inside a sandbox container.

computer-useagentsautomationanthropic

CrewAI

CrewAI is the fastest path to a working multi-agent prototype — role-based crews, linear workflows, readable code. LangGraph wins on production-grade control. Use CrewAI for content pipelines and research automation; LangGraph for anything with cycles, branching, or checkpointing.

crewaimulti-agentrole-basedcrews

Google Agent Development Kit (ADK)

Google's open-source Python SDK for building production-ready agents — code-first, A2A-native, deploys to Vertex AI Agent Engine. v1.0.0 went GA in 2025; v1.18.0 is the current release (November 2025).

google-adkgoogleagentsmulti-agent

LangChain

The base LLM application framework. Provides document loaders, text splitters, prompt templates, chains, and the LangChain Expression Language (LCEL) pipe operator for composing LLM pipelines.

langchainlcelllm-frameworkchains

LangGraph

LangGraph v1.0 is the production standard for stateful multi-agent orchestration in Python, offering fine-grained graph-based control with built-in checkpointing and human-in-the-loop support.

langgraphmulti-agentorchestrationpython

LangGraph Cloud / LangGraph Platform

LangGraph Cloud (now part of LangSmith Deployment) is the managed runtime for LangGraph agents. Handles persistence, horizontal scaling, and Studio UI. Cloud (SaaS) is fastest to start; self-hosted is available for compliance.

langgraphcloudplatformdeployment

LangMem

LangMem is LangChain's library for long-term agent memory — extracting, storing, and retrieving facts from conversations across sessions. Works with any storage backend. Integrates natively with LangGraph's memory store.

langmemmemorylong-termlangchain

Mastra

TypeScript-native agent and workflow framework from the Gatsby team — batteries-included alternative to LangGraph for JS/TS stacks, with built-in memory, evals, observability, and durable workflows.

mastratypescriptagentsworkflows

Multi-Agent Patterns

The three foundational multi-agent topologies (supervisor, swarm, parallel fan-out) and when each is worth the added complexity over a single agent.

multi-agentorchestrationsupervisorswarm

OpenAI Agents SDK

OpenAI's official Python framework for multi-agent systems (March 2025) — lightweight, model-driven handoffs, built-in guardrails; trade-off vs LangGraph is simplicity at the cost of persistence and HITL support.

openaiagents-sdkswarmhandoffs

Practical Agent Design

Single agent first — only escalate to multi-agent when tooling-rich single agents consistently fail. Tools have three types. Guardrails are layered, not singular.

agentsorchestrationtoolsguardrails

PydanticAI

Python-first agent framework by the Pydantic team — type-safe agents with Pydantic-validated outputs, dependency injection via RunContext, and automatic retry on validation failure.

pydantic-aiagentspythontyped-agents

ReAct Pattern

ReAct (Reasoning + Acting) is the foundational agent loop that interleaves chain-of-thought with tool calls and observations, grounding reasoning in actual tool results rather than hallucinated facts.

reactagent-looptool-usereasoning

Strands Agents SDK

Strands Agents is AWS's open-source Python SDK for building agents in a few lines of code. Uses the Bedrock Converse API under the hood; @tool decorator turns any Python function into an agent tool; native MCP server support. Used in production by Amazon Q Developer and AWS Glue.

strandsawsbedrockagents

Voice Agents

Voice agents run on either a pipeline (STT+LLM+TTS, 800ms–2s, most production) or a native multimodal model (GPT-4o Realtime/Gemini Live, 200–500ms); LiveKit Agents v1.5 is the leading open-source framework, VAPI/Retell for managed deployment.

voice-agentslivekitpipecatvapi