Bilingual

Beyond the Hype: Real-World MCP Support Across Major AI APIs
a day ago
- MCP (Model Context Protocol) is a promising development for connecting tools in an LLM-agnostic way, but real-world support varies significantly among major providers.
- Anthropic (Claude) offers seamless MCP integration out of the box, but its API is expensive and token consumption can be high for simple use cases.
- OpenAI provides remote MCP support but the author encountered a 'Session terminated' error and could not get it working, indicating it needs more testing.
- Google Gemini's MCP support is limited to enumerating local tools in JavaScript and Python SDKs, lacking true remote implementation and requiring client-side heavy lifting.
- There is a market opportunity for a provider hosting open-weight models (like Qwen3 or GPT-OSS) to deliver a polished, integrated MCP experience on hosted API endpoints.
Building Agentic Workflows in Python with LangGraph
8 days ago
- LangGraph represents an agent as a graph with state, nodes, and edges; state is a shared TypedDict, nodes are functions that update state, edges define execution order.
- MessagesState provides built-in message history with add_messages reducer to accumulate conversation automatically.
- Model calls are placed inside a node; tools are registered using @tool decorator and bound to the model via bind_tools.
- ToolNode handles tool execution; tools_condition routes graph between model and tool nodes in a ReAct loop.
- A checkpointer (e.g., InMemorySaver) persists conversation state across separate graph invocations using a thread_id.
- The standard ReAct pattern involves two model calls per tool use: one to decide the call, one to interpret the result.
Deprecated Accessor Footgun in AI SDK v7
21 days ago
- A bug in conversation history construction caused small models to fail multi-turn tool calling.
- The bug erased the model's own tool calls and results from its history, making follow-up tasks impossible.
- Correcting the history by including tool calls and results restored 100% success rates across all models.
- Frontier models masked the defect by performing well despite incomplete history, hiding the issue.
- Prompt mitigations interacted unpredictably with the broken history, sometimes causing catastrophic failures.
- Auditing conversation histories and adding regression tests are essential to avoid similar bugs.
- Whole-artifact rewrite designs are immune to this failure class as they have no hidden tool channel.
- Published benchmark results were corrected after identifying the bug, showing tools and rewrite at parity.

tool calling

Beyond the Hype: Real-World MCP Support Across Major AI APIs

Building Agentic Workflows in Python with LangGraph

Deprecated Accessor Footgun in AI SDK v7