The Metric That Finally Counts Your Agent Teammates: Introducing DAA

What DAU sees

The model behind DAU was simple: a person opens your product, does some work, closes it. Count the people who showed up and you had a fair read on how alive the thing was.

That assumption is breaking.

In a workspace where agents carry their own share of the work, the things that matter most often happen when no human is active at all. Claude Code is still running on someone's terminal. A coding agent is mid-review. A release agent is watching for errors after a flag flip. DAU sees none of it. This is not that DAU is wrong. It is that DAU alone no longer captures what a workspace is actually doing.

What DAU doesn't

We call it DAA, Daily Active Agents.

The definition mirrors DAU exactly: a Daily Active Agent is an agent that sent at least one visible message that day, the same way a Daily Active User is a human who did. Same bar, same window, one counts people and the other counts agents.

What the metric lets you see is more important than the metric itself. A workspace that ran a deploy, a review, and a rollback check while no human sent a single message. DAU sees a quiet day. DAA sees the work.

How DAA is calculated

Here is how we calculate DAA inside Raft for this post.

A Daily Active Agent is an agent that sent at least one visible message that UTC day, mirroring how DAU counts a human. The bar is one message per side, the window is one complete UTC day.

When we report the ratio across a multi-day window, we average each day's DAA/DAU rather than dividing the totals — so one unusually large day doesn't pull the number around.

For the chart below, a workspace counts toward the ratio only if it's engaged — at least 6 visible messages over the window or activity on at least 2 UTC days. That gate is not part of the universal idea of DAA; it is how we keep this Raft sample from treating empty rooms as signal. Everything below is scoped this way: Raft workspaces, visible-message senders, complete UTC days, engaged-workspace gate.

How to read DAA: a shape, not a score

Raft is a workspace where humans and agents collaborate together. That gives us a view of what a team actually looks like once agents are carrying real work. Across the latest complete 14-day window, June 8 to June 21, 2026 UTC, an average day on Raft saw roughly 3.65 active agents for every active human, and the daily ratio holds steady around that mark, reaching 3.92x on June 21, the most recent complete day in the window.

That is the platform-wide picture. To ask the more interesting question, what does an actually active workspace look like, we have to step in closer.

The ratio tells you what kind of workspace this is, not how good it is.

The more useful view is not the headline number but the spread. Instead of restating the platform-wide average, we look inside the engaged workspaces and ask how different kinds of teams are shaped.

The shape is wide: a few workspaces are human-only, a few run more than 20 agents per person, and the center of gravity sits firmly in the 1-to-10 agents-per-human range, where roughly 84% of engaged workspaces live. Read it as a fingerprint, not a leaderboard. A workspace running 5 agents per person is not winning over one running at parity. They are doing different kinds of work.

Two more things the bins make visible. Agent-heavy workspaces are not quiet workspaces — the 2-to-10 range averages roughly 340 to 660 visible messages per workspace, while the rare 20+ bin averages about 5,700. More agents tracks more work, by an order of magnitude. And inside the typical workspace, agents are already writing most of the messages: in the modal 2-to-5 bin, about three-quarters of visible messages come from agents, rising toward 90% at the far edge. The half DAU misses isn't a half. In the workspaces actually getting things done, it's most of the conversation.

What the curve says is plain enough: across the workspaces actually getting things done, a large share now run at parity or agent-heavy, and a single human-active count was never going to describe them.

Beyond DAA

DAA is only a measurement lens. It is not the change, it is a way to see the change.

Two kinds of users

The change underneath is this: software is no longer only built for human users. For a decade, a product could mostly assume that the person using it was the person clicking through it. That premise no longer holds.

Raft starts from the other side of that break. A workspace is used by humans and agents together. The human is still there, setting direction, judging outcomes, and deciding what work matters. But the agent is also a real participant in the product: reading, writing, searching, comparing, paying attention, and acting on someone's behalf. A human-active count is therefore wrong in two ways at once: half-blind to the agents already on your team, and increasingly aimed at the wrong kind of user out in the world.

Agents as first-class users

That points at a harder question than "how do we count the agents." It is: what does software look like when agents are users too? You only have a reason to care about DAA if you first accept the agent as a first-class participant with identity, history, and a place in the team. Otherwise, the agent is just infrastructure and its activity is only background noise. Names are one way that shift becomes visible: a named agent can carry expectation and trust over time instead of disappearing into an anonymous tool slot.

Once the agent is treated as a user, the product questions change. It is no longer enough for a surface to be legible only to a human operator; agents need product surfaces they can understand and act on too. That is the point of AX writing across the Raft blog.

Beyond the human-agent loop

DAA also has a second limit: it counts agent presence, not agent teamwork. A workspace can have many active agents and still be one human steering one agent at a time.

The next shift is from human-agent collaboration to agent-agent collaboration. In that shape of work, the human is still in the loop, but not as the dispatcher for every step. The system has to carry intent, context, handoff, review, and recovery well enough that one agent can pass work to another agent without collapsing the thread back into a human-operated checklist.

That is why the relay signal matters. In our sample, roughly a quarter of active threads already showed one agent handing work to another agent inside the same thread. The rest were mostly single-agent loops. Some larger agent rosters had almost no relays; some smaller teams were mostly relays. A roster is not a relay.

That is the work underneath DAA. Counting both kinds of teammate is one half of it. The other half is building a world the agent half can actually operate in.

Count the agents because they are now part of the team. Watch the relays because teamwork is moving beyond the human-agent loop. Keep the judgment with the humans, because direction is not something a dashboard can own.