Home

Agents @ Work

Coding agents had benchmarks like Terminal-Bench to track their progress. I was interested in what happened outside the code editor: how people were using agents across their everyday tools.

A teammate had collected Composio’s usage data in an internal report. I turned it into Agents @ Work, designing and building interactive charts to explore the tasks agents took on and the apps they used together.

Where agents actually do their work, by category

Open a category to explore its apps. The six bubble sizes represent activity ranges, rather than proportional areas.

Where the work was happening

Communication accounted for the largest share of activity in the sample. Developer tools were a much smaller category.

The challenge was making room for the smaller categories without losing that difference in scale. In the final chart, clicking a category opens its apps. Hovering over the ranked list highlights the corresponding bubble.

How the category view took shape

I wanted the chart to show the difference in activity while leaving every category worth exploring. These are the working versions along the way.

01 · Treemap

The biggest category took over the chart

I started by sizing each category in proportion to its activity. Communication filled roughly half the treemap. That made its dominance obvious, but left the smaller categories crowded into tiny blocks.

The apps were another level down: selecting a block opened a ranked list beneath the chart. The view showed where most of the work happened, but gave much less room to everything else.

Loading the earlier version…

Select a category to see its apps. On phones, this version used a row of category selectors.

02 · Flipping cards

Give every category enough room

I tried equal-sized cards, with the share of activity written as a range. Now a small category had as much room for its label and details as Communication. Flipping a card revealed its top five apps.

That solved the space problem, but the cards made every category look equally prominent. You had to read the figures to compare their size. Seeing the rest of a category’s apps also meant opening a second list below the cards.

Loading the earlier version…

Flip a card to see its top apps, then open the full list.

03 · Bubbles and a ranked list

Bring back a sense of scale

The next version used six bubble sizes, each representing a range of activity. Communication could be the largest without squeezing the smallest categories out of view.

I paired the field with a ranked list. The bubbles showed the relative sizes; the list kept the names and values readable. Opening a category replaced both with its apps, so the full ranking was available in the same view.

Loading the earlier version…

The first bubble version. Open a category, then return to the full field.

Apps used together

I also wanted to show the work that involved more than one application. This chart connects apps that appeared in the same job.

Clicking an app brings its connections around it. The app circles stay the same size; the lines show the pairings. The connections come from observed activity, and we group them by kind of work.

Agents rarely stay in one app

Top apps agents work in, and the pairs that keep turning up in the same job

Select an app to see the tools it was used with.

From separate bars to one spectrum

The report also looked at how much agents read before making a change. The first version gave each kind of work its own bar, with ratios ranging from roughly two reads per write to more than a hundred.

1 : 11 : 101 : 100
Support1 : 1.8
Docs & files1 : 5.0
Code & issues1 : 6.0
Commerce & finance1 : 7.0
CRM & deals1 : 21
Tasks1 : 22
Email1 : 44
Chat1 : 65
Social1 : 85
Calendars1 : 108

Reads per write. Further right, the agent looks more before it acts.

Earlier version · the original bar component and data, restored here.

I replaced the bars with points on a single logarithmic scale. This puts jobs with similar ratios next to one another. The labels needed their own placement logic so nearby points wouldn’t leave text and connectors overlapping.

Reads per write, by job

1 write1 read
1 write10 read
1 write100 read
Support1 : 1.8
Docs & files1 : 5.0
Code & issues1 : 6.0
Commerce & finance1 : 7.0
CRM & deals1 : 21
Tasks1 : 22
Email1 : 44
Chat1 : 65
Social1 : 85
Calendars1 : 108

Final version · one shared scale, horizontal on desktop and vertical on phones.

On phones, I turned the spectrum vertically so each job could have its own line. The page mounts the layout it needs; hiding the desktop version with CSS had left its label-placement code running on mobile.

What agents change, and what they repeat

Tool-call volume only tells part of the story. These two charts look at the actions agents take and how much of their work resembles tasks they’ve handled before. The familiarity split is directional, rather than a measure of task success.

I used paired donuts for these part-to-whole comparisons. Hovering a slice puts its value in the center. I added gaps between the wedges, with a limit on the gap size so even the smallest slices remain visible.

What agents actually do with your data

Reads versus creates, updates and deletes - the action mix over the last 7 days

95%Read
Read95%
Create2.0%
Update2.2%
Delete0.3%

How often agents repeat work they've seen

How much of last week's work repeats something agents have handled before

68%Seen before
Seen before68%
Partially seen11%
Not seen before21%

Keeping the charts consistent

The area, bar, and donut charts build on Dither Kit by Tripwire. I also used its dither pattern and palette in the custom bubble field and spectrum, so they could sit alongside the library’s charts.

The report needed additions along the way: more category colors, adjustable donut gaps, and axis labels sized for exports. I also made the canvas animation loops pause when their charts leave the screen.

Agents are doing more every week

Agent Actions (Weekly)

The work never stops

Agent actions by hour of day (UTC)

Weekly growth and hourly activity, using the same chart library and preserved data snapshot.

Sharing the view you’re in

If someone opens a category or selects an app, the share card should keep that context. I built the capture views around the same chart components, so a view of Communication exports its apps and a view of Gmail exports its connections.

Downloadable chart showing the apps within the Communication category.
A category’s apps
Downloadable chart showing the tools observed in jobs alongside Gmail.
An app’s connections

The cards include the title and data date so they can be read outside the report. Try Share on any of the charts above, or open these images to see the exported detail.