Architecture AI Updates: November 8, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. ByteByteGo Maps Where Six Agent Frameworks Put Orchestration, Lists Six Cost Levers and Traces a Claude Code Request
ByteByteGo’s EP232 system design newsletter, published November 7, compares how six frameworks orchestrate tasks, failures, shared memory and context across agents. LangGraph represents workflows as a graph whose nodes do work and whose edges define state transitions; CrewAI assigns each agent a role and a goal and runs tasks sequentially or under a manager agent; Microsoft Agent Framework targets organizational workflows with graphs that add built-in branching, checkpoints and pauses for human intervention; the OpenAI Agents SDK leaves it to the model at runtime whether an agent hands control to another agent or keeps executing; Google ADK is a hybrid in which the developer defines the strategy and a root agent splits and distributes tasks to sub-agents; and the Claude Agent SDK has no separate orchestration layer, since orchestration lives inside one agent loop that gathers context, calls tools and evaluates results until no tool calls remain. The comparison amounts to a choice of where control flow lives (an explicit graph, a manager or root agent, or the model’s own loop), which in turn decides where checkpoints, approvals and retries can be attached.
The same issue lists six ways to lower the cost of an AI system: prompt caching of stable prefixes such as system prompts and tool schemas, model routing that sends simple queries to a small model and hard ones to a frontier model, provider batch APIs that discount asynchronous jobs, a semantic cache keyed on query embeddings that reuses stored answers, trimming conversation history instead of resending all of it, and smaller models produced by distillation or quantization. It also traces one request (“Fix the failing test in auth.test.ts”) through Claude Code in eight steps, based on the Claude Code source: the CLI, IDE or SDK wraps the prompt with repository and file context, the agent loop proposes an action such as an edit to specific lines, the permission system checks it against rules, the approved action is dispatched as a tool call that runs in a shell, cloud or sandbox environment, the result returns to the loop, and the turn is persisted to state and streamed back, repeating until the model stops asking for tools. Source