AI agents
AI coding agents in late 2026: from one assistant to a team
Claude Code, Cursor and Google Antigravity have all moved the same way this year: from one chat in your editor to several agents working in parallel.
The shift
A year ago the typical setup was one AI assistant in one editor, answering one prompt at a time. By late 2026 the three tools I use every day, Claude Code, Cursor and Antigravity, have all moved to the same model: you describe the work, and a coordinator splits it across several agents running in parallel, often in the background.
Claude Code
Anthropic describes Claude Code as "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools." It runs in the terminal, IDEs, a desktop app and the web.
- Subagents: spawn multiple agents that "work on different parts of a task simultaneously."
- Hooks: run your own shell commands before or after the agent acts, for example a formatter or a test gate.
- MCP for connecting to external tools, plus background agents and an Agent SDK for building your own.
Cursor
Cursor's September Projects release introduced coordinators that run sub-tasks "as many in parallel as the work needs." An Agent SDK followed in October, and in November Cursor added voice chat with agents from its Agents Window.
Google Antigravity
Google's Antigravity app is built to "orchestrate multiple autonomous agents working in parallel across independent projects." Its feature list reads like a summary of where the whole category is going:
- Dynamic subagents, defined and created on the fly for a task.
- Scheduled tasks that run on a cron, plus Artifacts for reviewing what agents produced.
- Customisation through Skills, MCP servers and JSON hooks, and a CLI for managing background subagents.
What changes for us
The job moves from typing code to directing and reviewing it. In my own work at Inventive Rhino that means three habits:
- Write the task down properly first. Agents running in parallel need a clear goal and clear boundaries, or they collide.
- Keep the guard rails in code, not in prompts: tests, type checks and lint as hooks or CI gates.
- Review every diff. Faster output is only useful if it is correct, accessible and secure.