This archive accompanies Manual Code Review Is Optional and the agent-authored analysis about building Retirement Lab by specifying work and operating software agents rather than manually writing source code. It exposes selected primary artifacts from the private repository without publishing the product wholesale.
Behind these examples is an analysis of 1,077 commits, 1,016 merged PRs, and 711 agent runs. The records below make the workflow inspectable; the aggregate rates come from the full corpus. See the snapshot corrections for the original report’s earlier commit count.
Examples from the Essay
| Record | What it shows |
|---|---|
| Database connections held while waiting | A reviewer finding, the merged connection-lifetime correction and a regression test |
| Other independent review findings | Consent ordering and duplicate simulations, with recorded findings and corrections |
| Testing layout rules | The original overlap instruction, committed test and later extensions |
| A currency error and its repair | The withdrawal and repair PRs behind the essay’s consequential failure example |
Four Complete Workflow Records
| Record | What it shows | Outcome |
|---|---|---|
| A full spec that passed directly | A detailed parallel-compute spec, focused implementation/test diff, one-pass gates, and independent review | 860 changed lines; no human intervention recorded; no known follow-up repair |
| A near-median task that passed directly | The smaller spec task path, including its complete instruction and end-to-end data/UI/test change | 239 changed lines versus 242 median for tasks; no human intervention recorded |
| Independent gates finding real gaps | One failed semantic-contract test and four P1 review findings over three review rejections | Corrected automatically over four implementation attempts; no human intervention recorded |
| A reviewed change that still shipped a bug | A difficult 3,670-line PR, 15 pre-merge findings, explicit operator intervention, and a later two-line product fix plus regression test | Strict bug repaired 35 days after origin |
What the archive supports
The records show the mechanism behind the quantitative claims: committed specs, isolated implementation runs, independent verification, a separate reviewer, retry loops, operator escalation, merged code, and later repair work. They also demonstrate that “autonomous” does not mean “first draft” and that extensive machine review does not eliminate escaped defects.
Data files
- Essay metrics and counting definitions
- Essay metrics (JSON)
- Aggregate summary (JSON)
- Phase statistics (CSV)
- Full analysis report (Markdown)
- Defect-classification rubric (Markdown)
- Selection and redaction manifest (JSON)
- Artifact checksums (JSON)
Read the methodology and limits before using the numbers as a comparison baseline.