Every feature, improvement, and fix — documented from day one. We ship fast and believe in transparency about what we're building.
When the agent is working on a multi-step task — writing files, running commands, installing packages — it now shows its reasoning between each action in real time. You'll see the same thinking panel that appears in regular chat: a live clock, the model's actual thoughts streaming in, and an expand/collapse toggle. Each step gets its own fresh reasoning, so you can follow exactly why the agent chose to do what it did next. Models with built-in reasoning (like Qwen3 and DeepSeek) produce visible thoughts; models without it simply skip the panel silently.
Previously, when the agent wrote a file — a README, a script, a document — nothing happened on screen for several seconds, then the entire file appeared at once. It felt like the app had frozen. Files now stream into the terminal block progressively at roughly 60 frames per second, so you can actually read the content as it's being written. A live line counter in the block header (e.g. '142 lines') shows you how far along the write is, then switches to the elapsed time once complete.
File write and file read blocks used to stay collapsed by default — you had to manually click them open to see what was happening inside. They now open automatically the moment the agent begins the action, the same way command blocks already did. This means you always see the agent's work as it happens without having to interact with the UI.