Open the Agent tab
The tab appears when the agent publishes an agent, and opens by itself. The agent publishes a draft as soon as it has written the agent’s definition, so you can read the graph before the proof runs. Each later publish opens the tab again with the new version. The tab keeps itself current. When the session’s machine wakes, it shows grey placeholder bars and reads the agent again once the machine is up. Close the tab with its ×; to bring it back, click + at the end of the tab strip and choose Agent.If you don’t see the Agent tab after an agent build, the agent hasn’t published yet, or the Agent builder skill is switched off for your sessions. Ask “Publish the agent”, or switch the skill back on. People you share the session with see the canvas, the details panel and the eval result, but not Try it or the eval setup, because only the owner can run the agent or change its evals.
Read the canvas
The canvas reads top to bottom, in the order the agent runs. Each kind of card keeps one colour everywhere: on the card, in its details and on the minimap.
The step whose result is the answer carries an
answer tag. Steps that don’t depend on each other share a row inside an outline labelled parallel · 2 (or however many). Nothing runs between them, so they can run at the same time; whether they do is up to the framework. Lines have no arrows, because top-down order gives the direction, and a dashed line is one the compiler inferred. An Inputs or Answer card with seven or more rows shows five and a Show all button, such as Show all 9.
The small dot that travels the lines is decoration. It shows how the agent is wired, never that it’s running. Real run state comes from Try it, which marks each card as its step runs.
Inspect a step or tool
Click a card and its details dock to the right of the canvas, which keeps the card in view. Its lines light up and the rest dim; in a lane, the other steps stay lit and the outline turns blue, because they run together. Close the panel with its ×, a click on empty canvas, or Esc.Step details
The Inputs and Answer cards open the same way, listing every input or field in full.
Tool details
Click a tool chip, on a card or in a step’s details, to open the tool rather than its card.Walk through the flow
The walk-through moves through the agent one card at a time, opening each card’s details. It’s the quickest way to explain an agent in a meeting. Click the play button (“Walk through the flow”). A row unfolds with the controls and your place, such as3 / 6 · parallel 1/2.
The arrows stop on every card, each lane step included. Play moves one row every 2.4 seconds, and a lane is one row. The row folds away 10 seconds after play reaches the answer. Clicking a card stops play.
Arrange and zoom
Drag any card to move it. This browser remembers the place for this session and agent. Moving a lane’s card removes its outline until you reset. Click Zoom to unfold the controls beside it; on a trackpad, two fingers pan and a pinch zooms.
The keys work once the canvas has focus, so click an empty part of it first. Zoom runs from 30% to 200%. Below 55%, cards show only their names, large. Agents with more than 12 cards get a minimap in the bottom-right corner. Every shortcut is on Keyboard shortcuts.
See how the agent did on its evals
Once an eval run finishes, the answer step’s card carries its pass bar, such as “10 of 12 passed” and a percentage: green for passes, red for failures, orange for attempts that couldn’t be graded. It always shows the latest finished run, and no other card shows eval results. The flask button sits beside Try it. Once a run has finished, its dot gives the verdict: green passed, red failed, amber anything else.
Viewers of a shared session see “Not tested yet: no eval run has finished” once evals exist but no run has finished. Before that, they see no flask button.
What if the canvas can’t draw?
Instead of a wrong picture, the canvas says what’s wrong and the one thing to do. Most fixes are a message to the agent.
The not-ready states also show the command that fixes them. You don’t need to run it yourself. Send this:
Next steps
Try your agent
Send it a real input and watch each card on the canvas light up as its step runs.
Set up evals
Turn the flask button green: written test cases, personas and a cost before anything runs.
Build an AI agent
What to say to get an agent, the ten frameworks, and the proof every agent passes.
Read results and fix failures
Open the run behind the pass bar and let Alpha propose one fix.