Compare agents on the same task
Launch the identical assignment across different CLIs and inspect their responses, verification, files, and diffs side by side.
All product areas
04 / Agent Tasks
One assignment. Multiple agents. You choose the result.
Run a focused task with one agent, or send the same prompt to several agents across different CLIs, models, and effort levels. Every attempt stays isolated and comparable until you decide which result belongs in the workspace.
Actual Agent Tasks capture
WHAT IT CHANGES
Less reconstruction.
More visible progress.
Choose what reaches the workspace
Each attempt remains separate until you select the strongest result and deliberately apply that candidate to the real project.
Choose the exact runtime
Select the CLI, model, and effort level using the same availability-aware controls as the desktop app.
Work in isolation
The run receives its own Git worktree so experiments do not leak into the working folder.
Review the response
Final agent messages, checks, changed files, and diffs are presented as a result instead of terminal debris.
Preview before apply
Open the candidate worktree in a terminal, file browser, side-by-side diff, and live browser before changing the workspace.
WHAT IS BUILT IN
The working surfaces behind Agent Tasks.
Availability-aware configuration
Sourceweave only enables installed, compatible CLIs and exposes model, provider, and effort controls when that adapter supports them.
Readable progress instead of a wall of logs
Agent responses, commands, file changes, attention states, and raw output are separated so the final message remains easy to find.
Put competing attempts side by side
Compare different agents by final response, model and effort, changed files, additions and deletions, verification, and the actual diff before selecting a winner.
Run the isolated result
Start the suggested development command, discover the local URL from terminal output, inspect files and media, and test the result in the embedded browser.
HOW IT WORKS
A deliberate path through Agent Tasks.
- 01
Define the assignment
Choose the workspace and describe the bounded change you want the agent to make.
- 02
Choose one or more agents
Configure the exact CLI and model for a single run or send the same request to several agents across different CLIs.
- 03
Follow live progress
Structured output separates commands, tools, thinking state, warnings, and the final response.
- 04
Compare, choose, and apply
Inspect each isolated branch and its verification evidence, compare the candidates, then apply only the result you trust.
KEEP EXPLORING