Agent frameworks

Agent frameworks#

What you’ll learn in this part

  • ReAct’s parser + tool-call loop from scratch.

  • Three strategies for structured LLM outputs (prompt-only, Pydantic validate+retry, FSM-constrained).

  • State-machine-shaped agents (LangGraph-style 50-line clone).

  • DSPy’s signature + MIPROv2 prompt optimisation.

  • The Model Context Protocol as a 2-file server + client.

  • Conversation-driven (AutoGen / AG2 / Microsoft Agent Framework) vs role-driven (CrewAI) multi-agent idioms.

  • Evaluating agents with τ-bench / SWE-bench-shaped benchmarks.

The primitives here (tools, state graphs, handoffs, guardrails) are the same building blocks used by OpenAI Agents SDK (handoff chains), Google ADK (supervisor hierarchies), Pydantic AI (dependency-injected agents), and smolagents (code-as-actions). The 2025 A2A protocol (Agent-to-Agent) adds horizontal discovery across agent boundaries; see the glossary for an overview. Microsoft Agent Framework 1.0 (April 2026) unified AutoGen 0.4 and Semantic Kernel into one production SDK; the patterns in notebook 06 apply directly to it.

Key terms used in this part#

  • ReAct is the baseline loop pattern used to teach tool-using agents.

  • structured outputs means constraining model outputs to a schema instead of free-form text.

  • FSM (Finite-State Machine) constraints are one way to enforce valid output structure during decoding. XGrammar v2 is the current production-grade FSM engine. It runs constrained decoding about 3× faster via its Structural Tag protocol.

  • MCP exposes tools/data over JSON-RPC so clients can use them uniformly.

  • DSPy frames prompts/pipelines as optimizable programs.

  • A2A (Agent-to-Agent Protocol, Google ADK 2025) is the open standard for cross-framework inter-agent delegation and discovery.

  • handoff and guardrail are the two new first-class primitives introduced by the OpenAI Agents SDK.

  • Microsoft Agent Framework is the April 2026 merger of AutoGen and Semantic Kernel into a single production-ready agent SDK.

Ecosystem snapshot (mid-2026)#

The agent framework landscape has consolidated around a few common patterns:

  • LangGraph v1.0 (stable): now the most-starred agent framework on GitHub, used in production by companies including Uber, LinkedIn, and Klarna. The 1.0 release adds durable state (automatic execution persistence), built-in human-in-the-loop approvals, native sandboxing, sub-agents, first-class MCP support, and a distributed runtime via the CLI. It passed CrewAI in GitHub stars during early 2026.

  • AutoGen / AG2: Microsoft’s AutoGen v0.4 ships streaming and event-driven architecture; the community maintains the proven v0.2 lineage as ag2ai/ag2 with typed tools and dependency injection. AG2 Beta (autogen.beta) is a ground-up redesign with multi-provider LLM support and first-class testing. Both run on the same core SelectorGroupChat / GroupChatManager API.

  • CrewAI (v1.12): added agent skills, hierarchical memory isolation, Qdrant Edge memory backend, and native support for OpenRouter, DeepSeek, Ollama, vLLM, Cerebras, and Dashscope providers.

  • Google ADK v1.0 (stable): hierarchical agent tree where a root agent delegates to sub-agents; ships stable releases in Python, Go, Java, and TypeScript. Introduces the A2A (Agent-to-Agent) protocol v1.0 (in production at 150+ organisations) so a LangGraph or CrewAI agent can be invoked by an ADK agent without bespoke adapters. Announced at Google Cloud Next 2026 alongside Workspace Studio (no-code agent builder) and managed MCP servers via Apigee.

  • Pydantic AI: a lightweight, type-safe alternative used for simple agents where full LangGraph state management is more than needed.

  • MCP (Model Context Protocol): the 2026-07-28 specification (release candidate; final text ships July 28, 2026) is the largest revision since launch — it removes the Mcp-Session-Id protocol session so any server instance can handle any request, drops the initialize/initialized handshake entirely, and rewrites authorization around standard OAuth/OIDC RFCs instead of bespoke wiring. A new extensions framework (reverse-DNS-namespaced, independently versioned) is how the Tasks extension (long-running async tool calls via tasks/get, tasks/update, tasks/cancel) and MCP Apps (server-rendered interactive UIs in sandboxed iframes) now ship. The Enterprise-Managed Authorization extension already reached stable status, adopted by Anthropic, Microsoft, and Okta. MCP is now the de-facto standard for AI tool integration; X (formerly Twitter) shipped a hosted MCP server for its platform API in July 2026. Enterprise adoption requires SSO-integrated auth, audit trails, and gateway behavior.

Reading order#

No mandatory prerequisites. CPU-only.

  1. 01_react_from_scratch: three-line agent loop, regex parser, and three tools.

  2. 02_structured_outputs_three_ways: flaky-LLM simulator; prompt vs validate+retry vs FSM.

  3. 03_langgraph_state_machines: StateGraph clone with conditional edges.

  4. 04_dspy_3_miprov2: 3×3 (instruction, demo) grid; MIPROv2 as 5-sample random search.

  5. 05_mcp_server_client: JSON-RPC 2.0 tool server and synchronous client.

  6. 06_autogen_0_4_vs_crewai: draft/critique/revise pipeline two ways (AutoGen/AG2 and CrewAI).

  7. 07_agent_evaluation_suite: success rate, trajectory efficiency, and code-patch success.