Claude Code: Opus 4.7 and Agentic Orchestration
Claude Code has evolved into a powerful terminal-native agent with the release of v2.1. The most significant update is the optimization for Opus 4.7 and the introduction of the /ultrareview command, which launches a parallel, multi-agent cloud review for branches and GitHub PRs.
The tool has also expanded its agentic capabilities with "Agent Teams" (research preview), allowing coordination between a team lead and teammates via shared task lists. For power users, new commands like /loop (recurring intervals) and /effort (reasoning budget control) provide granular control over the agent's behavior. Additionally, the introduction of the /remote-control feature now allows users to bridge their terminal sessions to a browser or mobile device.
Read more — Claude Fast Blog
Google Gemini Code Assist: "Finish Changes" and Outlines
Google has aggressively updated Gemini Code Assist in early 2026, introducing a "show, don't tell" pair programming experience. The new "Finish Changes" feature observes in-progress modifications to complete tasks autonomously, while "Outlines" generates high-level English summaries interleaved within the source code for better comprehension.
On the infrastructure side, the Gemini CLI now includes GitHub Actions for autonomous issue triage. Paid tiers (AI Pro/Ultra) have also seen their context windows expand to 1 million tokens, allowing the AI to process roughly 30,000 lines of code in a single prompt. The introduction of "Agent Mode" in VS Code and IntelliJ now leverages the Model Context Protocol (MCP) to connect to external services.
Read more — Google Developers Blog
Andrej Karpathy: From Coding to "Manifesting"
Andrej Karpathy has sparked a conversation about the fundamental shift in software engineering, moving from "writing code" to "orchestrating agents." In February 2026, he released microgpt, a zero-dependency GPT implementation in ~240 lines of Python, designed to teach the algorithmic essence of LLMs.
More significantly, his AutoResearch project (or "The Karpathy Loop") demonstrates the power of agentic iteration. By using a loop where an agent forms a hypothesis, modifies code, and validates against a metric, the system successfully optimized NanoGPT parameters beyond Karpathy's own manual tuning. He argues that developers are entering an era of "manifesting" their will through agent strategies, suggesting that the primary target for documentation should now be agents rather than humans.
Read more — The AI Corner
