The peek-a-boo agent bar
- Role
- Design Direction
- Team
- Serene Chen
- Timeframe
- 2025–2026
- Outcome
- Shipped as Workato's persistent agent bar; patent filing planned
Keeping several agents visible without stealing the screen
Running several AI agents at once creates a different problem: how do you supervise all of them without being tied to a single screen. We wanted someone to be able to:
- Start several agents on different tasks
- Leave those tasks running
- Continue working manually elsewhere in Workato
- See when an agent finished or needed help
- Respond without losing the work already in progress
Why this is particularly difficult in Workato
A Workato user can belong to several workspaces — perhaps one for Sales, another for Legal, and another for Finance. Each workspace can contain development, test, and production environments. An agent might therefore be running in a different workspace, environment, and product area from the page the user is currently viewing. If agent activity lived inside a page-level chat panel, people would have to remember where each agent had started and navigate back there to check on it. That would turn parallel work into a monitoring task. We needed an agent layer that belonged to the entire product — not to a single page, workspace, or environment.
The first version tied agents to the wrong place
An early concept placed an agent-status indicator on the environment switcher. I pushed back on it in review. The switcher answered one question: “Which environment am I viewing?” Agent status answered another: “Does any active work need my attention?” Combining them made both harder to understand. More importantly, the concept still treated an agent as belonging to the user's current environment.
Prototyping the principle
I used our internal product scaffolding and Claude Code to build a rough version of the idea I wanted to discuss with the team. The prototype was not intended to be the final design. It made three principles tangible:
- Running agents should remain visible throughout the product.
- Their status should not depend on the current workspace or environment.
- Anything that needs a decision should be reachable in one action.
We tried the top of the screen first
Serene and I explored several placements. One concept used a centered notch, similar to the top of a MacBook screen. It could expand just enough to offer a peek at active agents. We liked the sense that agent activity was always nearby without becoming a permanent panel. But it collided with an existing control in the recipe editor, where people switch between editing and testing. We then tried a full-width bar across the top. It created more room for active agents, but expanding it pushed the product down and competed with navigation that already lived there. Both concepts failed for the same reason: the top of the product was already doing too much.
The breakthrough was moving it to the bottom
Serene placed the interaction at the bottom of the viewport. The move resolved several problems at once:
- It avoided the product's existing navigation and mode controls.
- It could appear on any page without restructuring that page.
- It remained in peripheral view without competing with the current task.
- Expanding it created a natural transition into a detailed agent view.
- The action for starting a new agent could sit near the conversation that followed.
The shipped pattern
In its resting state, the bar provides a compact view of active agent sessions. It can surface:
- How many agents are running
- Which agents need input
- What completed while the user was elsewhere
- A shortcut to open any session
- A way to start another agent
Collapsed to full-screen, on demand
People can move among pages, workspaces, and environments without losing track of the agents working for them. When more context is needed, the bar expands into a richer view or opens the complete agent experience. The transition happens only when the user asks for it.
The agent bar
My contribution: I reframed the problem as a product-wide information-architecture issue, rejected the environment-bound direction, established the interaction principles, and built the first prototype to communicate them. Serene led the detailed interaction design and found the bottom placement that made the pattern work.

