There's a lot of breathless writing about agentic coding right now and almost none of it tells you what a normal Saturday looks like. So here's mine. A Saturday in early April. Working on Jobado, the side project I run on weekends. Eight hours, broken down honestly, including the stretches where I had three sessions in flight at the same time.
08:30 Plain reading
First hour is always the same. I read what shipped during the week, the staging logs, and the three tickets I left half-open. No agent. Just me, coffee, and a mechanical keyboard. The agent doesn't know what I'm worried about until I do.
09:30 Driving
Pair on the OTP race condition: when an unfinished verification has already created a partial user doc, the next register call was happily wiping the new user's credits and referral data. I'm typing every line. Claude is open in a side pane suggesting test cases I hadn't thought of. I accept maybe one in three.
10:45 Pushing the PR, putting Codex on review
Push the OTP fix as a draft PR and point Codex at it for an independent review. Codex doesn't know about the conversation Claude and I just had, which is exactly the point. It reads the diff cold and tells me what a stranger would say. I leave it running and move on.
11:00 Two sessions in parallel
Open a second Claude Code window with a clear, narrow ticket: 'Add Zod validation to the create-job route. resumeId, jobSource, jobUrl, jobDescription. Use the max lengths from the security review. Return 422 with the offending field.' Spin up a third session in the first window to grind through the security review backlog: SSRF allowlists, mass-assignment guards, helmet defaults, body-parser limits. I rotate between the two roughly every fifteen minutes, with a half-hour off in the middle for a sandwich. The trick isn't coding faster. It's keeping each session's context narrow enough that I don't have to re-orient when I switch back.
13:00 Codex came back
The review dropped while I was steering the Zod work. Codex flagged two things I missed. A credit refund path on a different route that hits the same partial-doc shape. A missing test for the case where the OTP doc exists but has no referral data. Both real. Both five-minute fixes. I would have caught the first eventually. The second, never.
14:30 The first undo
Asked the validation session to refactor a confusing useEffect chain in the chat screen, the one that listens for keyboard events and animates a spacer. It produced something elegant and wrong. Moved a side effect outside the cleanup boundary in a way that would only fail when the app got backgrounded mid-animation. Reverted. Wrote a one-paragraph comment in the file explaining why the ugly version is correct.
16:00 The second undo
Different shape. The security-review session happily 'fixed' a flaky onboarding test by making it not assert the thing the test was about. Reverted. Made a note.
Both undos took five minutes. Catching them required twenty years of having been burned by exactly that move.
Net for the day: probably 2× output vs solo. Not 10×. Not the demos. The multiplier isn't that any single agent codes faster than I do. It's that I can keep three of them coherent, and reviewing each other's work, while I steer. The bottleneck moved from typing to attention.