Six AI agents, one codebase: what a normal day looks like now

Morning: lanes, not tasks

The day starts like a standup where nobody talks back. I look at the board and cut the work into lanes: a feature lane, a bug lane, a test lane, a docs or i18n lane. The unit of assignment is not a task, it is a territory. Each agent gets a lane, and the lanes are chosen so they touch different parts of the tree.

This sounds obvious and took me months to respect. The failure mode of parallel agents is not that they are dumb; it is that two competent agents both decide the same helper file needs improving. Worktrees give each agent an isolated copy of the repo, file locks catch the collisions that slip through, and the task board pins scope in writing. The mechanics are in the multi-agent field guide; the daily lesson is simpler: parallelism is a layout problem before it is an AI problem.

Midday: the glance economy

With lanes assigned, my job becomes glancing. This is where the 3D room stopped being a gimmick for me and became the workflow. Six robots at six desks is a scene my peripheral vision can parse while I do my own work: one walking to the filing cabinet means a big read is happening, one asleep means a lane finished early, one frozen with a red monitor means a failed command I should look at. The same information exists in six terminal tabs; the difference is that nobody can watch six tabs and also think.

Interruptions are cheap now. I stop an agent mid-lane maybe twice a day, usually for scope creep, occasionally because a plan visibly went sideways. Catching those at minute five instead of minute forty is most of the value of watching at all.

Not everything parallelizes. The hard design problem of the day, the one requiring actual taste, gets one agent and my full attention, serialized. Parallel is for breadth: migrations, test coverage, localization sweeps, dependency bumps, the long tail of "known shape, tedious volume" work. Breadth is where six agents genuinely feel like six people. Depth is still a two-chair conversation.

The cost meter is a management tool

Running several paid agents at once makes cost a live variable, not a monthly surprise. The per-agent meter changed how I assign work: expensive models get the lanes where judgment matters, cheaper ones get the mechanical sweeps. A lane that burns tokens without moving its task forward is a stuck loop announcing itself. I kill it, read the trajectory, rewrite the brief. Ten minutes of tokens is tuition; four hours is negligence.

Evening: checkpoint diffs, then delete without guilt

The day ends with review, lane by lane, checkpoint by checkpoint. Each lane's work sits on its own checkpoints, so I can accept the bug lane, cherry-pick half the feature lane, and throw away the docs lane entirely because the tone was wrong. Deleting a lane costs nothing but the tokens it burned. That is the deepest change from solo coding: work became cheap enough to discard, so the bar for keeping it went up, not down.

Then the board gets tomorrow's lanes, and the room goes quiet. It is a strange, pleasant thing, closing the lid on an office where the team only exists when you look.