Frameworks & Tools
LangChain, LangGraph, LlamaIndex, DSPy, multi-agent frameworks, coding agents — and how to choose.
- lessons
- 6
- steps
- 53
- time
- 75 min
LangChain and LangGraph
Compose pipelines with LCEL pipes, keep dependencies lean with the package tiers, and move to LangGraph when the agent needs cycles, state and checkpoints.
- What LCEL pipes give you over hand-written glue
- Which LangChain packages belong in production
- Why agents need cycles, and how LangGraph models them
- Checkpoints, interrupts and time travel per thread
LlamaIndex and DSPy
Two specialists: LlamaIndex for data-shaped work (parsing, indexing, event-driven workflows) and DSPy for turning prompts into compiled, optimisable programs.
- LlamaIndex's data-first model: Nodes, parsers and retrievers
- Event-driven Workflows and data-driven fan-out
- When to pair LlamaIndex with LangGraph
- DSPy: signatures, modules and optimizers
Multi-Agent Frameworks
Semantic Kernel and AutoGen merged into Microsoft Agent Framework, CrewAI's crews and flows, handoffs, typed agents in Pydantic AI and Mastra, and how to stop runaway loops.
- Why new Microsoft-stack work starts on Agent Framework
- Semantic Kernel's plugin model and why planners were removed
- CrewAI roles, processes and Flows
- What a handoff forwards, and what it drops
Observability with LangSmith
Trace every node of an agent run, turn feedback into datasets, score runs with judges, compare variants, and budget for tracing itself.
- The observability pyramid: value, flow, raw traces
- Trace attribution: find the node that really failed
- Datasets and judges as unit tests for LLM apps
- Cheap decision-model judges vs LLM judges
Coding Agents
How Claude Code and its peers work: the agent loop and tools, CLAUDE.md, subagents, permission modes, safe CI pipelines, cost, and the wider coding-agent landscape.
- The coding-agent loop: tools gated by a permission layer
- CLAUDE.md as project memory, reviewed like code
- Why auto mode is a convenience, not a boundary
- A safe headless pipeline in CI
Choosing a Framework
Match the framework to the job, keep vendor SDKs at the leaves, learn the primitives every framework converges on, and survive the churn with pins, lockfiles and evals.
- Pick a framework from the job it has to do
- Raw SDK, thin layer or framework
- Neutral orchestrator at the core, lab SDKs at the leaves
- Why tutorials break, and the silent changes that don't
Guide chapters · Frameworks & Tools
- LangChain Deep Dive7 min
- LangGraph Orchestration5 min
- LangSmith Observability6 min
- LlamaIndex6 min
- DSPy: Programming Language Models6 min
- Semantic Kernel3 min
- Microsoft Agent Framework, CrewAI, and the Agent SDK Landscape17 min
- Framework Selection Guide11 min
- Claude Code: The Autonomous Coding Agent19 min
- OpenCoder: AI Coding Agents Landscape24 min
- Pydantic AI and Mastra: Typed Agent Frameworks (2026)12 min
- Navigating Framework Churn21 min