It’s interesting to see Anthropic lean so heavily into a CLI-first approach for agentic coding. We’ve seen a lot of success with tools like Aider and various IDE extensions, but a first-party tool that can natively leverage the model’s specific tool-use strengths feels like the right direction for reducing friction.
The real challenge with these agents is usually the "context ping-pong"—constantly switching between the editor, the docs, and the terminal to verify changes. By living directly in the shell, this starts to feel less like a chat interface and more like a specialized terminal multiplexer that actually understands the project state. It’ll be worth watching how they balance autonomy with safety as the tool evolves.
I've noticed in using Opus 4.6 that it seems more proactive, whereas Sonnet 4.6 stops carefully after each task. Opus is definitely smarter, while Sonnet seems to follow instructions more literally, and sometimes fails to take all instructions into account. I'm using spec kit and wedow/ticket (beads alternative) in OpenCode with Opus / Sonnet through a Copilot subscription.
It’s interesting to see Anthropic lean so heavily into a CLI-first approach for agentic coding. We’ve seen a lot of success with tools like Aider and various IDE extensions, but a first-party tool that can natively leverage the model’s specific tool-use strengths feels like the right direction for reducing friction.
The real challenge with these agents is usually the "context ping-pong"—constantly switching between the editor, the docs, and the terminal to verify changes. By living directly in the shell, this starts to feel less like a chat interface and more like a specialized terminal multiplexer that actually understands the project state. It’ll be worth watching how they balance autonomy with safety as the tool evolves.
I've noticed in using Opus 4.6 that it seems more proactive, whereas Sonnet 4.6 stops carefully after each task. Opus is definitely smarter, while Sonnet seems to follow instructions more literally, and sometimes fails to take all instructions into account. I'm using spec kit and wedow/ticket (beads alternative) in OpenCode with Opus / Sonnet through a Copilot subscription.
>resistant to correction
Is this emergent behaviour or coded behaviour?
That report is really law quality. Claude Code didn't get credentials to all those platforms by default...