Do AI coding agents make you faster? Measure it before you scale it

The Friday retro is where it happens. You say the thing you know is true: since the agents arrived, you have been shipping like never before. Then someone scrolls the merge log for the quarter, and your count of finished, surviving changes is about what it was in March. The feeling of speed and the fact of speed, it turns out, have quietly stopped agreeing.

This is not a you problem. In the sharpest experiment run on this question so far, METR timed experienced open-source developers working real issues in their own repos, randomized between AI-allowed and AI-forbidden. With AI, tasks took 19 percent longer. Afterward, the same developers estimated the AI had made them about 20 percent faster. That is a forty-point gap between the feeling and the stopwatch, in the people most likely to know better.

Maybe your agents do make you faster. Plenty of teams measure real gains. The point is that you do not get to know by feel, and the hours and dollars you are about to spend scaling up deserve a better instrument than vibes. The instrument is a ledger with two columns.

Why you cannot feel agent speed

Delegation feels productive even when it is not. The hour spent steering an agent registers as leverage. The redo on Thursday registers as a fresh task, not as the true cost of Tuesday. The gains arrive labeled and the costs arrive anonymous, so the books cook themselves.

Agent work also slices your attention in ways memory flattens. Kicking off a run, checking on it, nudging it, re-reading the diff: each slice is small, none of it feels like the task, and all of it was the task. Feel is the wrong tool here. Counting is the right one.

Count shipped tasks, not lines of code

The credit column has one honest unit: a finished task, merged and still alive a week later. Not lines generated, not suggestions accepted, not pull requests opened, not tokens burned. Those are activity metrics, and agents produce activity on demand.

For the unit to mean anything, tasks need to be roughly comparable, which is one more reason to cut work into agent-sized tasks: a board full of half-day tasks makes a countable week.

DORA's research on AI keeps arriving at the same shape: AI amplifies the delivery process it lands in, and throughput that outruns review shows up later as instability. Count what merged, and count what stayed merged. A credit column alone proves nothing.

The debit column: review, rework, babysitting

Speed claims die in the debit column. Four lines to keep:

• Review minutes. Agent diffs still cost human reading time, and review is the new bottleneck. Write down what it actually costs you, per task.
• Rework. Any agent change you reopened inside the week: a bounce back, a follow-up fix, a quiet rewrite by hand.
• Interventions. Every time you stepped in to unstick, redirect, or answer a question the brief should have answered.
• Abandoned runs. Sessions that produced nothing you kept. They cost real time and leave no artifact to remind you they happened.

None of this says agents are a bad deal. It says the deal has two sides, and you have only been reading one of them aloud.

Read the ledger like a diagnosis

A week of honest numbers does more than answer "faster or not." Each oversized debit points at its own fix:

• Review minutes dominate: your tasks are too big. Smaller cards produce smaller diffs, and small diffs are the difference between reviewing and rubber-stamping.
• Rework dominates: your briefs are loose and nothing forces the agent to prove its work. Tests the agent must satisfy turn "looks done" into "is done" before it ever reaches you.
• Interventions dominate: the task was underspecified, or the repo fights cold readers. Fix the brief before blaming the model.

And if the ledger comes out clearly positive: congratulations, you have just written the business case for running more agents in parallel, with your own numbers instead of a vendor's.

Where measuring goes wrong

The number becomes the goal. The moment "tasks shipped" turns into a target, tickets split like cells and the count doubles while nothing changes. Goodhart's law does not spare you because the workers are agents. Keep the ledger a private diagnostic, not a report card.

The ledger becomes a hobby. Standing dashboards have a standing cost. One honest week, repeated after each workflow change, beats a permanent measurement habit that never ships anything itself.

One week is treated as truth. A single developer's week is noisy: one gnarly bug moves every number. Read direction, not decimals, and rerun before concluding much.

"Not faster yet" is read as a verdict. METR's developers were weeks into these tools, and skill with agents compounds for months. A negative week is a workflow finding that tells you which debit to attack, not proof the whole idea is a mirage.

A board that keeps the ledger for you

The reason nobody keeps this ledger is that it lives in six places: scattered terminals, a git log, your memory of how often you stepped in. The fix is to run agent work somewhere that records it as a side effect.

SanuDesk is a desktop app built around that idea. Work starts as cards on a Kanban board; deploy a card and it becomes an agent session in a tiled grid, Claude Code, Codex and Gemini side by side. The board then keeps your books without being asked. The done column is your credit column. Cards waiting in review are your review debit, and every card you bounce back with feedback is a rework tally you did not have to remember. The grid makes interventions visible too: stepping in stops being an invisible habit when it happens on one screen. Recurring chores run as Loops with a journaled run history, so routine work gets counted without you keeping score. It is bring-your-own-model, running on the agent subscriptions you already pay for, and the board and grid are in the free plan (details on pricing).

Run the one-week ledger

Start Monday. Before anything else, write down your prediction: how many tasks you expect to ship this week, and how much time you think review will take. Sealed, dated, no edits.

Then tag every task agent-led or by-hand, and keep four tallies: tasks shipped, tasks reopened, review minutes, interventions. On Friday, put the totals next to Monday's prediction.

Two findings fall out. The totals tell you whether the agents are paying and which debit to fix first. The gap between prediction and totals tells you how far to trust your own sense of speed, which METR suggests may be the most expensive number you have never checked.

Download SanuDesk free and let the board keep the ledger, or see how the grid, the board and Loops fit together on the features page.