Agentic coding stacks

Coding agents are being reorganized into teams: one assistant plans, others search the repo, run commands, write code, and hand work back. (docs.langchain.com) The basic idea is simple: instead of one chatbot holding an entire codebase and task list in one context window, developers split work into smaller roles with separate memory and tools. LangGraph documents fixed workflows and dynamic agents for this pattern, and its multi-agent packages include “supervisor” and “swarm” setups for routing work between specialists. (docs.langchain.com, reference.langchain.com) CrewAI markets the same structure as “agents, crews, and flows,” with guardrails, memory, knowledge, and observability built in for production use. Microsoft’s AutoGen still describes itself as an event-driven framework for scalable multi-agent systems, but its GitHub README now says the project is in maintenance mode and points new users to Microsoft Agent Framework. (docs.crewai.com, microsoft.github.io, github.com) The shift is also showing up in coding products, not just developer frameworks. Cursor says its agent can search a codebase with semantic and agentic search, and its January 2026 release added subagents that run in parallel with their own prompts, tools, and models. (cursor.com, cursor.com) Anthropic is pushing the same model in Claude Code. Its product page says Claude Code can read a codebase, make multi-file edits, run tests, and deliver committed code, while Anthropic’s recent materials describe subagents, hooks, and “teams of Claude Code Sessions” that divide parallel work. (anthropic.com, anthropic.com, resources.anthropic.com) Open-source tools are following the same path. OpenCode describes itself as an open-source coding agent for the terminal, and its site says users can start multiple agents in parallel on the same project; Agently says its framework turns multi-step agent behavior into orchestrated systems and now ships installable skills for coding agents such as Codex and Claude Code. (opencode.ai, opencode.ai, agently.tech, agently.tech) What changed over the last year is that “agentic coding” stopped meaning a single autocomplete box with better prompts. Current docs from Cursor, Anthropic, LangGraph, CrewAI, and OpenCode all describe coordination features — subagents, flows, supervisors, cloud agents, hooks, tracing, or shared workspaces — that treat software work as a sequence of delegated jobs. (cursor.com, docs.langchain.com, docs.crewai.com, anthropic.com, opencode.ai) That structure fits the way programming teams already work. One agent can inspect a large repo, another can run shell commands, another can browse docs or internal tools through Model Context Protocol, and a lead agent can merge results instead of re-deriving everything from scratch. (cursor.com, anthropic.com, docs.langchain.com) The tradeoff is that more agents also mean more moving parts. LangGraph emphasizes persistence, debugging, and human-in-the-loop controls for long-running workflows, CrewAI sells observability and monitoring, and Anthropic frames hooks and permissions as a way to keep autonomous coding systems inside guardrails. (docs.langchain.com, docs.crewai.com, anthropic.com) The result is a new coding stack taking shape in 2026: not one chat window that sometimes writes code, but an orchestrated workspace where multiple assistants share a repo, split tasks, and report back. (cursor.com, anthropic.com, docs.crewai.com)