Too many agents, too many windows
AI agents let me do more at once. Keeping track of them became another task.
I’m one developer, often working with several AI agents at once: one refactoring a module, another chasing a failing test, a third drafting a feature plan.
AI coding agents let me take on more work at once. But the faster they moved, the harder it became to keep up with them.
Where it broke down
An agent would stop to ask a question in a tab I wasn’t looking at. Another would finish while I was watching a deployment. I would return to a terminal and try to remember which task belonged there.
Getting a second model to review a result meant copying it into another app and explaining the project again. A UI bug meant finding the screenshot and dragging it into the right conversation.
Every time I lost the thread, work slowed down. I needed a clearer view of the projects and agents I was directing.
Every project in one place
In Workbench, each project has its own named terminal sessions and agent conversations. I split panes to suit the task, and Workbench restores the layout and working folders when I reopen it.
The terminals run my login shell, so I use the Git, SSH and package-manager setup I already have. That makes the workspace fit the way I develop.
Know which agent needs an answer
Agent status is visible in the workspace and menu bar panel. I can see which sessions are working, waiting for me or finished, and go to the conversation that needs attention. Notification support depends on the tool and my settings.
That changes how I manage the work. Instead of repeatedly checking every tab, I can focus on one task and return to another when it needs a decision.
I also use agents to help watch Azure DevOps deployments. The agent needs the appropriate tools and access, and I still decide whether to approve a stage. Workbench helps me keep track of that conversation alongside the rest; the deployment work happens through the tools available to the agent.
Bring another model into the conversation
I use Codex, Claude Code and Grok for different tasks. Workbench connects through their installed command-line tools and my own accounts.
When I switch providers within a conversation, Workbench prepares a handoff summary for me to review. The next model receives that context so I can ask it to assess the previous work. A summary can miss details, so I check that the important decisions and files are included.
Steer the work as it develops
Plan mode gives me a place to explore an approach before implementation. Build mode lets the agent make changes using the permissions supported by its integration.
Questions and approvals appear in the conversation, and the activity view helps me inspect what the agent ran. The available controls, mid-task steering and detail of change reports vary by provider.
I can also queue follow-up instructions while a task is running, then edit or reorder them before they are sent. That fits how I work: the next decision often comes to mind before the current task has finished.
Show what I mean
Screenshots and attachments help me explain a UI problem without describing every pixel. Workbench saves captures into the project so I can bring them into the right conversation.
For design discussions, I can ask an agent for an interactive HTML mockup and review it in the conversation. It gives me something concrete to click through and revise with the agent. Mermaid diagrams serve a similar purpose when discussing architecture or a flow.
What changes for one developer
I still have to choose priorities, review results and make decisions. Workbench makes it easier to see where those decisions are needed.
I can open my laptop, find the agents waiting for me and get their work moving again. The agents bring the speed; the workspace helps me stay in charge of the work.
The final post looks at the other part of that job: keeping the project’s decisions and plans available when a session ends or another agent takes over.
Explore Workbench or download Workbench.
Workbench requires macOS 27 or later on Apple Silicon. Each AI tool is installed and signed in separately. Permissions, notifications and steering capabilities depend on the provider integration.