David Vekslergithub.com/DavidVeksler
7,537commits in 2026, so far

One developer, a fleet of AI agents, and 10,000 commits by mid-November.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
appears on of this year's commits.
$git log --all --since=2026-01-01 --date=short --format=%ad | sort | uniq -c

Then the agents clocked in.

$git log --all --date=format:%Y --format=%ad | sort | uniq -c

A decade of commits, then a wall.

$git log --all --date=format:%H:%M --format=%ad | sort | uniq -c

I sleep. They ship.

$git log --all --format='%(trailers:key=Co-Authored-By,valueonly)' | sort | uniq -c

Who co-wrote the code.

Claude Code signs its work with a Co-Authored-By trailer. Codex and Copilot sometimes do. Whatever carries no trailer lands in the grey bucket, which also holds the commits I typed myself.

$git log --all --format='%(trailers:key=Co-Authored-By,valueonly)' | grep -o 'Claude [A-Za-z]* [0-9.]*'

Every new model goes straight to work.

Weekly Claude-signed commits by model generation. The workflow is model-agnostic: runbooks and scripts stay put, the model underneath gets swapped as soon as a better one ships.

$find ~/.claude/projects -name '*.jsonl' | xargs jq -c 'select(.type=="assistant") | .message.content[] | select(.type=="tool_use") | .name'

One prompt, 44 actions.

$for r in */; do git -C "$r" rev-list --all --count --since=2026-01-01; done

90 repos touched this year.

$git log --all --since=2026-01-01 --name-only --format= | sed -E 's/.*\.//' | sort | uniq -c | sort -rn

14 languages in one year.

$git log --all --since=2026-01-01 --format=%ad --date=format:'%F %H' | sort | uniq -c | sort -rn | head -1

Personal records.

$cat ~/AGENTS.md

How one person ships 100 commits a day.

The volume comes from a small set of rules every agent reads before it touches a repo. They are written down, versioned, and enforced by machines where possible.

  1. Runbooks are the spec

    Every repo opens with an AGENTS.md routing table: which doc governs which task. Agents read the smallest path and stop guessing.

  2. Scripts before tokens

    Anything done twice becomes a script. Agents orchestrate scripts instead of re-deriving what code can compute, which keeps runs fast and cheap.

  3. The night shift

    Recurring judgment work runs as scheduled routines between 1 and 5 AM, each with hard limits, per-run caps and a fail-closed rule.

  4. Git is the safety net

    Agents commit and push without asking, in small commits. History is the backup and the audit log. Deployment is the one human approval gate.

  5. Rules become gates

    A rule an agent might forget becomes a pre-push hook or a build-failing lint. Machines enforce it every time.

  6. Right model, right job

    Haiku for bulk mechanical work, Sonnet for volume, Opus for judgment, Fable for specs that cheaper models then implement.

Case studies GitHub profile