Agent code authorship

An open research project measuring what repository history can—and cannot—tell us about code written with coding agents.

The study combines public agent-provenance signals with Sourcegraph's indexed Git history. It follows known agent-attributed lines after they land in established open-source repositories, while keeping unsupported prevalence and quality claims explicitly unavailable.

Read the investigation

The self-contained article and its script-free figures are available at results/agent-code-authorship-sourcegraph.html. It can be downloaded and opened directly in any modern browser.

Current result

The current population share is not identified. Public provenance signals find known positives, but they do not supply the contemporary human denominator needed to estimate how much of all open-source code is agent-written.

Durability is identifiable in a separate frozen cohort:

• 656,070 agent-attributed Go and Python lines across 96 repositories
• 90.2% estimated to survive at 365 days
• 84.9% estimated to remain unchanged at 365 days
• separate estimates for Claude Code, Codex, GitHub Copilot, and Cursor

These numbers describe known agent-attributed code after it lands. They are not an estimate of agent-written code's share of open source, and deletion is not a revert or a quality judgment.

Why prevalence remains unavailable

Commit trailers and other preserved provenance establish high-confidence agent examples, but most assisted work is unlabeled. Treating unlabeled modern code as human would build the answer into the control group.

The repository includes experiments with stylistic classifiers and era-adjusted controls. Their identification gates reject a population estimate: a model can distinguish old from new code without learning authorship, and the available modern controls do not resolve that ambiguity. The published article reports the boundary instead of substituting a style proxy.

Repository map

Path	Contents
authorship/	Collection, validation, estimation, and rendering code
study/	Frozen protocols, schemas, manifests, and study inputs
results/	Compact result artifacts, figures, and the rendered article
corpora/	Compressed feature records used by the earlier classifier work
data/	Cohort definitions and control-policy evidence
tests/	Unit, integration, artifact-contract, and browser tests

Large raw Sourcegraph responses, local clones, caches, and agent-workspace metadata are intentionally excluded. The checked-in manifests and checksums document the inputs used by the published artifacts.

Reproduce the published article

Python 3.12 is the reference runtime.

python3 -m venv .venv
. .venv/bin/activate
python3 -m pip install -r requirements.txt
python3 -m authorship.build_v3_blog_post

The renderer uses the frozen JSON artifacts in study/ and results/. If a neighboring Sourcegraph checkout contains the blog fonts, they are embedded; otherwise the output remains self-contained and uses system font fallbacks.

Rebuild the standalone study figures with:

python3 -m authorship.build_report_figures

Validate

Install the development dependencies and run the Python suite:

python3 -m pip install -r requirements-dev.txt
python3 -m pytest -q
python3 -m ruff check authorship tests

The browser checks cover responsive layout, navigation, runtime errors, and WCAG A/AA violations:

npm install
npx playwright install chromium
npm run test:e2e

Method at a glance

• Agent attribution requires explicit, auditable provenance; ambiguous cases remain unlabeled.
• Survival follows lines from their introducing commit through later repository states using indexed blame and Git history.
• Estimates are repository-balanced so a handful of very large repositories do not define the answer.
• Bootstrap intervals resample repositories, preserving codebase-level clustering.
• Underpowered or mixed-evidence strata are reported as not identified.
• Cohorts, revisions, schemas, and analysis gates are frozen before reporting.

For artifact-level detail, see results/SOURCEGRAPH_STUDY_ASSETS.md and the protocols under study/.

Scope and interpretation

The repository studies established public repositories and known agent-attributed code. It does not measure private code, unlabeled assistance, developer productivity, defect rates, or the causal effect of any coding agent.

Prevalence, survival, reverts, and quality are different estimands. Do not add them, average them, or treat one as a proxy for another.

License

Released under the MIT License.