Chat with your agent. Spin up tasks. Keep the trace.
Agent chat is a working conversation with your engineering agent - not a toy chatbot. Ask questions, attach context from your inventory, and promote any reply into a real, staged task without leaving the thread.
Conversation that knows your repo.
Sessions are scoped to a workspace. The agent can read files, search the repo, run shell commands, and create staged tasks - every action is logged in the same activity trace.
Why does the sidebar test button look misaligned in dense mode? I'm seeing it sit ~2px lower than the other rows.
Looking at Sidebar.tsx and the dense token in DESIGN_TOKENS.md - row height is 26px in dense, but the test button's icon-tile has margin-top: 2px hardcoded. Removing that and switching to align-self: center matches the other rows exactly.
Good catch. Let's fix it as a task - patch only that file, with a snapshot test.
Task created. The planner will draft a 2-step plan; I'll wait for your approval before workers run.
- Scope
- packages/frontend/src/components/Sidebar.tsx
- Workers
- 1 · isolated worktree
- Engine
- Studio Engine
- Gate
- Approve plan before run
Pull from your inventory in two clicks.
The Attach menu surfaces your workspace context inventory inline - markdown docs, uploads, saved URLs, and selected directories. Drop in anything the agent should consult; the trace records the exact items used.
- Pick from Context item or From computer
- Attached items appear as chips above the composer
- Hash-tracked - re-run with the same context any time
- Per-message scoping - only what's attached is in scope
Want a deeper dive on the inventory itself? See Context inventory.
From conversation to task in one click.
Talks. Doesn't ship.
- Replies but can't run anything on your repo
- Context is whatever you remember to paste
- No way to promote a reply into work
- No trace, no evidence, no audit
Conversation wired to the pipeline.
- Repo-aware: reads files, searches, runs commands
- Attach context items from the workspace inventory
- Create a staged task from any reply
- Every tool call lands in the run trace
Try a real conversation with your repo.
Agent Studio is in private alpha. Drop your email for a build and setup guide.