Three checks on the multiplier · follow-up to Two Eras of Shipping

Under the multiplier.

96%of my Tano PRs were never reviewed by another person. A bot reviewed 98% of them, and I merged them myself.PostHog: 19% had no human review

~1.8×more of my PRs needed a follow-up fix from me within a week.about 14% at Tano vs 8% at PostHog

1.6×more PRs open at once, while each one moved through 7 times faster.2.9 vs 1.85 open · 10 h vs 70 h each

Check 1 · the learning curve

Output grew fivefold inside Tano while merge time stayed flat

My merged PRs per week at Tano. The company, the codebase and I stay the same across the chart, so this is the cleanest view of skill with agents. The markers are tooling changes from the repo's history. Hover a week for its count.

Swipe sideways to see the whole chart.

Month	PRs merged	Median lines added	Median time to merge

What this shows

From February to August my weekly output grew about five times and my median PR doubled in size. Median time to merge stayed under an hour in every month, including the first. So the fast merges came with how Tano works from day one. The growth in volume is the part I learned.

Check 1 · where the speed comes from

I finish each PR faster. I only juggle a little more at once

Little's law ties the three figures together: PRs in flight = PRs merged per day × days each PR stays open. Each row compares PostHog with Tano on its own scale.

PostHogTano

How many agent sessions run at once

Share of the time any Claude Code session in the Tano repo was active, split by how many ran at the same time. From the 513 session logs Claude Code still keeps (June–September 2026).

Check 2 · who reviews the work

Review moved from teammates to a bot

Share of my merged PRs. Reviews and comments I wrote on my own PRs are not counted.

PostHog · 859 PRsTano · 1,638 PRs

What this shows

At PostHog a teammate reviewed 82% of my PRs, usually within about 2 hours of opening. At Tano an AI reviewer reads every PR, from April onwards without exception, and I merge after its review. Review threads per PR looked flat in the first infographic because the bot's comments replaced my teammates' comments.

Check 3 · rework

More of my PRs need a second pass

Share of my PRs followed within 7 days by one of my own fix PRs to the same files. The bars show the estimate after I checked a sample of matches by hand; the lines show its 95% range.

How the estimate narrows	PostHog	Tano
A later fix PR touches any of the same source fileswithin 7 days; shared hub files excluded	13.3%	35.7%
The fix PR sits mostly inside the original's filesat least half of the fix's files	10.8%	22.0%
Matches that are real rework when read by hand40 random pairs per company	70%	63%
Estimated PRs needing a follow-up fix	7.6%	13.8%

PostHog · a real matchfeat(flags): Allow disabling flags on first load↓ 4 hours laterfix(flags): Make sure flags are reloaded only once on identify calls

Tano · a real matchfeat(partner-agent): agentic graph invoker, delete v1 one-shot executor↓ 3.5 hours laterfix(partner-agent): stop task-completion hallucination, unblock retries

What this shows

The revert rate in the first infographic fell from 2.3% to 0.4%, but reverts undercount rework: at Tano I fix forward instead of reverting. Counted as follow-up fixes, rework per PR roughly doubled. Rework per line of code fell, because Tano PRs are five times larger.

What the checks say

Four things I learned

Rework rose while reverts fell.

Test share held and reverts fell, but follow-up fixes per PR rose about 1.8 times.

The merge speed is mostly a review model.

A bot reviews and I merge. That was true from my first month, before most of the growth in volume.

The volume is the learned part.

Weekly output rose fivefold inside one company and one codebase, alongside worktrees, repo skills, AI review and per-directory AGENTS.md files.

Parallel agents matter less than you might expect.*

PRs in flight rose only from 1.85 to 2.9. Half of my active agent time is a single session.

Method and caveats

Reviews: from the GitHub API for the same 859 PostHog and 1,638 Tano PRs. Bots are accounts GitHub marks as bots plus posthog-bot. The Tano AI reviewer is Claude; a few PRs also had reviews from Codex and Cubic. A bot counts as a reviewer when it left a review, an inline comment or its verdict comment.
Rework: only fix PRs I wrote count, so fixes by teammates are missed at both companies. Test, docs and generated files are ignored, and so are hub files touched by more than 1.5% of PRs. Hand labels are my reading of the two titles; 40 pairs per company gives the ranges shown.
Tano has about 11 times more of my PRs per week, which makes accidental file overlap more likely. The strict match and the hand check exist to remove that effect; some may remain.
Little's law figures use each PR's open time, from creation to merge, averaged over each tenure. Agent sessions are active when a log entry follows the previous one within 10 minutes. The logs cover June–September 2026 only, because Claude Code deletes older ones.
Milestones come from the repo's git history. Model releases are not marked, and they probably account for part of the curve.