How much of GitHub's most-starred AI tooling is written by coding agents?
The full git history of 52 of the most-starred repositories in GitHub's AI-tooling topics (324,088 commits), checked for AI signatures, churn and what each project publishes about its own results.
What the data shows
AI-signed commits are common
In the median repository, 16.9% of the commits in its last 180 days of activity carry an AI signature (13.6% over its whole history). 21 of 52 are at 25% or more and 10 at 50% or more. 38 sign at least 5% of their commits, 9 more commit instructions for coding agents while signing less, and 5 show neither signal (unknown, not “no AI”).
AI share does not track popularity
Stars and AI-signed share show no detectable link (Spearman ρ = -0.05, p = 0.72). The ten most-starred repos range from 0% to 73.6% AI-signed.
More fix commits, no more discarded code, at the same age
At day 180, these repos label 2.0 times as many commits as fixes (34.1% vs 17.3%), have lower churn (1.92x vs 3.77x) and keep more of the code they write (67.1% vs 41.7%), against 12 famous pre-Copilot projects.
AI use and recency cannot be told apart here
The fix share, churn and survival gaps are concentrated in the 31 projects started in 2025 or later: among them, repositories that sign at least 5% of their commits and those that sign less show no detectable difference, and the 10 repos started earlier do not differ detectably from the baseline. Heavy AI use is concentrated in the newest projects (AI share vs start date: ρ = 0.72), so this sample cannot separate AI use from recency. At one year (20 repos old enough), the fix share gap remains but the churn and survival gaps are no longer detectable.
Many publish their own results, including unfavourable ones
24 repos publish a benchmark harness with committed results; 15 of those measure whether the tool does its job (tokens saved, accuracy, cost, task success) and 11 are AI-signed. Several publish numbers that went against them. 3 publish nothing quantitative.
How much AI is in the commits
How the 52 repos sort
Signed means at least 5% of commits carry an AI co-author trailer, an AI agent as author or committer, or a “Generated with” footer, over the whole history or the last 180 days of activity. 44 of 52 also commit agent instructions (CLAUDE.md, AGENTS.md and similar) and 7 run a coding agent in CI.
AI-signed share of recent commits, by repo
Number of repos by the share of their last 180 days of activity that is AI-signed.
Highest AI-signed share, last 180 days of activity
Which tools sign the commits
85.3% of the 36,892 signed commits are Claude's. Claude Code adds a trailer by default, so tools that sign less often are under-counted.
Projects that switched recently (6)
Last 180 days of activity at least 15 points more AI-signed than the whole history.
Sign under 5% but commit agent instructions (9)
| Repository | Stars | AI-signed | Agent configuration in the repository |
|---|---|---|---|
| deepseek-harness | 276,014 | 0% | .agents/, .claude/, AGENTS.md, CLAUDE.md |
| hermes-agent | 261,582 | 2.9% | AGENTS.md |
| ollama | 186,319 | 0.1% | AGENTS.md, CLAUDE.md |
| langchain | 149,190 | 0.3% | AGENTS.md |
| ragflow | 93,749 | 2.4% | .agents/, .github/copilot-instructions.md, AGENTS.md |
| OmniRoute | 85,531 | 3% | .github/copilot-instructions.md, AGENTS.md, CLAUDE.md, GEMINI.md, .github/workflows/claude.yml |
| herdr | 49,219 | 0.3% | .agents/, AGENTS.md |
| elephant-goldfish-model | 14,254 | 0.7% | .agents/, .gemini/, AGENTS.md |
| gentle-ai | 8,586 | 0.3% | .claude/, AGENTS.md |
The remaining 5 (anything-llm, PageIndex, QwenPaw, conductor, open-webui) sign under 5% and commit no agent configuration. Their AI use is unknown.
Popularity
Spearman ρ between stars and AI-signed share: -0.15 over the whole history (p = 0.30) and -0.05 for the last 180 days of activity (p = 0.72). Stars and fix share: ρ = -0.0 (p = 0.99). More-starred repos churn more (ρ = 0.33, p = 0.017).
The ten most-starred
| Repository | Stars | AI-signed, all time | Last 180 days | AI evidence |
|---|---|---|---|---|
| superpowers | 314,290 | 13.3% | 6.1% | signed |
| ECC | 299,342 | 8.1% | 7.9% | signed |
| deepseek-harness | 276,014 | 0% | 0% | agent config, <5% signed |
| hermes-agent | 261,582 | 2.9% | 2.8% | agent config, <5% signed |
| firecrawl | 199,609 | 13.1% | 42.1% | signed |
| AutoGPT | 187,937 | 20.4% | 73.6% | signed |
| ponytail | 186,540 | 24.1% | 24.1% | signed |
| ollama | 186,319 | 0.1% | 0.2% | agent config, <5% signed |
| transformers | 168,646 | 1% | 10% | signed |
| dify | 160,649 | 5.3% | 6.7% | signed |
Fixing and churn at the same age
Churn depends on a project's age, so each repository is checked out at its last mainline commit 180 and 365 days after its first commit and measured the same way. The baseline is 12 famous projects (over 40k stars) whose first year ended before GitHub Copilot became generally available (2022-06-21). At HEAD, the median repo here has a fix share of 36.3%, a churn ratio of 2.08x and 65% survival.
Medians at day 180: 41 of these repos vs 12 baseline projects.
These repos vs the baseline, at day 180 and day 365
| Median | These repos, day 180 | Baseline, day 180 | These repos, day 365 | Baseline, day 365 | ||
|---|---|---|---|---|---|---|
| Commits | 1318 | 678.5 | p = 0.082 | 2013.5 | 1119.5 | p = 0.045 |
| Fix % | 34.1 | 17.3 | p < 0.001 | 27.3 | 17.1 | p < 0.001 |
| Churn ratio | 1.92 | 3.77 | p = 0.008 | 2.47 | 3.56 | p = 0.41 |
| Survival % | 67.1 | 41.7 | p = 0.011 | 57.55 | 43.65 | p = 0.46 |
| Repositories | 41 | 12 | 20 | 12 |
p-values: two-sided Mann-Whitney U, 10,000 permutations, not corrected for the number of tests. The baseline shows 0% AI-signed commits at both ages, which checks the detector's false-positive rate. At day 180, 41 of the 52 repos are compared: 8 have less than 180 days of history (deepseek-harness, elephant-goldfish-model, get-shit-done, gsd-2, holaOS, munder-difflin, ponytail, thinkrail) and 3 snapshots predate Copilot. Day 180 is the main comparison: at day 365 only 20 repos qualify, and they were far less AI-heavy at that age (median 0.8% AI-signed).
Is it AI, or is it when the project started?
Within these repos, AI share goes with fix share and not clearly with the other two: fix share ρ = 0.46 (p = 0.003); churn ρ = -0.22 (p = 0.17); survival ρ = 0.21 (p = 0.18). But AI share also rises with how recently a project started (ρ = 0.72, p < 0.001): 23 of the 31 projects started in 2025 or later signed at least 5% of their commits by day 180, against 0 of the 10 started earlier.
| Median, day 180 | Started 2025+, signed ≥5% | Started 2025+, signed <5% | Started 2022–2024 | Baseline | 2025+: signed vs not | 2025+ vs baseline | 2022–2024 vs baseline |
|---|---|---|---|---|---|---|---|
| Fix share | 36.2 | 34.65 | 22 | 17.3 | p = 0.47 | p < 0.001 | p = 0.18 |
| Churn | 1.89 | 1.86 | 2.79 | 3.77 | p = 0.57 | p = 0.003 | p = 0.37 |
| Survival | 69.2 | 69.25 | 52.95 | 41.7 | p = 0.60 | p = 0.003 | p = 0.38 |
| Repositories | 23 | 8 | 10 | 12 |
The fix share, churn and survival gaps are concentrated in the 31 projects started in 2025 or later: among them, repositories that sign at least 5% of their commits and those that sign less show no detectable difference, and the 10 repos started earlier do not differ detectably from the baseline. Heavy AI use is concentrated in the newest projects (AI share vs start date: ρ = 0.72), so this sample cannot separate AI use from recency. A less-signed project is not necessarily one without agents, so the comparison inside the newest projects is between more and less visible AI use.
What the projects publish about their own results
Strongest evidence each repo publishes about its own performance
Of the 24 repos with a harness and committed results, 15 measure whether the tool does its job (tokens saved, accuracy, cost, task success) and 9 measure only the speed of the software itself (AutoGPT, Paradev, agent-orchestrator, cmux, deepseek-harness, happy, orca, transformers, yao).
Several publish results that did not favour them, among them ECC, archify, context-mode and caveman. 3 repos (agents, conductor, elephant-goldfish-model) publish no quantitative results at all.
Results are the projects' own claims, located by file and line in the audited commit and classified by evidence type. None was re-run.
The 15 outcome benchmarks, AI-signed repos first
Click a row for the summary of what the repo publishes and links to the exact files.
caveman82.3% AI-signed
Caveman publishes three results with a runnable harness and committed data: a 10-prompt, single-run output snapshot on claude-opus-5-5 in which, on top of an 'Answer concisely.' control, `/caveman` cuts output a further 3% at the median and `/ultracave` 35%; the caveman-compress fixtures (6,198 → 4,138 tokens across five memory files); and a browser-snapshot benchmark (121 vs 15,704 tokens on a 200-row table, while on a small form caveman's full result is 2.34x larger than Playwright's). The headline input-token result (885,793 → 591,673 tokens, 33.2%, 18 of 18 answers right, with the HTML case 9.9% worse) and the Auto result (43.7% vs 44.3% resolved at 0.46x the cost, an interim run on 85 of 169 tasks) are committed as written reports: docs/WRAP-BENCHMARK.md states the first cannot be independently reproduced from this checkout, and docs/AUTO-BENCHMARK.md states the raw trajectories are not in the repository. The 63.6% (Elastic), 1.4 to 2.4x cost cut (Adobe Research) and p = 0.82 (JetBrains) figures are third-party results, and docs/HONEST-NUMBERS.md lists output reduction against default verbose replies as 'Not published' and states that no quality claim is published because no reviewed fidelity snapshot is committed.
Evidence: README.md:116, README.md:247, README.md:246
academic-research-skills70.4% AI-signed
academic-research-skills commits held-out measurement suites with preregistrations, run records and raw outputs: on seeded-defect manuscripts the reviewer panel reached recall 1.00 on both defective fixtures with 0 clean-control false findings while severity agreement fell from 0.672 to 0.607 (and from 0.663 to 0.536 after the #581 batch), and a preregistered audit put the panel's false-claim rate between 12/805 (1.5%) and 104/805 (12.9%). A guard prompt lowered judged claim-strength drift from 7/16 to 1/16 without eliminating it, the Wording-Pattern Advisory's held-out miss rate went from 0.375/0.344 to 0.094 with 0/16 false fires (judge protocol, no runner script), and the gpt-5.6-sol citation-verifier bakeoff recorded recall 0.90 vs 0.80 on 10 non-existent references (p95 latency 25.0 s vs 49.6 s) from committed receipts and an offline scorer. Smaller checks are reported as session counts, including a routing fix that passed 3 of 6 sessions and did not meet its preregistered acceptance rule; the per-mode token and cost table in docs/PERFORMANCE.md is an estimate measured on Opus 4.x, and several release notes mark prompt-level changes as unmeasured.
Evidence: evals/heldout/reviewer_seeded_defects/README.md:287-291, evals/heldout/reviewer_seeded_defects/README.md:295-296, evals/heldout/panel_claim_factuality/PREREGISTRATION.md:10-14
claude-mem66.9% AI-signed
The read-gate eval (scripts/eval-read-gate.ts, `npm run eval:read-gate`) has a runnable harness and committed per-case tables: with the gate on, mean cost per run on a 49 KB file was $0.077 vs $0.124 (-38%) on Sonnet and $0.171 vs $0.243 (-30%) on Opus, and -1% and 0% on a 19 KB file, where an earlier version of the gate had cost Opus 20% more. The work-view SQLite benchmark (scripts/benchmark-work-view.ts) has its synthetic-fixture and approximately 3 GB real-database latencies summarized in docs/development/viewer-sync-recovery.md; the real database is not in the repository, raw reports from both benchmarks go to a gitignored reports/ directory, and the read-gate eval needs paid model calls. The Smart Explore comparison (17.8x fewer tokens for discovery, 19.4x for targeted reads) is a written A/B report with no script in the repository, and the README and CHANGELOG figures (~10x, 90%, 200x) are stated without data.
Evidence: tests/evals/read-gate/README.md:87, tests/evals/read-gate/README.md:41, docs/development/viewer-sync-recovery.md:126
graphify52.7% AI-signed
graphify's BENCHMARKS.md reports results against named baselines with cost and judge-agreement figures (90.6% agreement, Cohen's kappa 0.81): LOCOMO recall@10 0.497 vs BM25 0.362 and mem0 0.048, LongMemEval-S 76% tied with dense RAG, and ERPNext key-fact coverage 82.0% vs 70.8% for a grep-and-read baseline (n=6); on LOCOMO QA accuracy supermemory scores higher (49.7% vs 45.3%) at about 11x the ingest cost ($15.67 vs $1.40). The harness these results point to (memory/runner.py, crosstool/run.py) is not in the repository, and no per-run outputs or spend ledgers are committed. The 71.5x token-reduction figure is backed by in-repo code (graphify benchmark) and a committed graph.json, with corpus tokens estimated from word counts rather than measured. The Graphify Cloud claim of cutting agent token spend by 80% is stated without data.
Evidence: worked/karpathy-repos/review.md:11-28, BENCHMARKS.md:19, BENCHMARKS.md:20
gsd-core48% AI-signed
gsd-core publishes no results on task success, output quality or end-to-end agent cost. Its measured numbers are static prompt-size reductions from two offline token-count scripts with committed JSON baselines: with the opt-in `workflow.compact_content` mode (off by default), the six spine/detail workflow splits come to 15.5% fewer tokens (113,676 to 96,055) and the 32 compact variant files to 21.91% fewer (122,467 to 95,637), counted with gpt-tokenizer as a stand-in for Claude's tokenizer and without a check that output quality holds; both rerun deterministically with `npm run benchmark:compact-content` and `npm run benchmark:compact-content-variants`. The '~40%' static-overhead reduction from prompt thinning and the '~75%' smaller command files are stated without a measurement.
Evidence: tests/fixtures/compact-content-benchmark-baseline.json:3, tests/fixtures/compact-content-variant-benchmark-baseline.json:3, docs/features/context-window-aware-prompt-thinning.md:7
headroom32.3% AI-signed
Headroom's main token-savings table (21–57% fewer input tokens on four seeded scenarios, 178,024 → 104,261 tokens in total) is reproducible offline: benchmarks/index_proof_table.py and its committed output benchmarks/results/index_proof_table.txt give the same numbers. The latency, accuracy (GSM8K 0.870 → 0.870; TruthfulQA 0.530 → 0.560, which the README calls no detectable difference), HTML extraction (F1 0.919), tool-search and Vertex AI (76.7% fewer prompt tokens) results have harness code with results recorded in markdown; no raw run files are committed for them, and several need paid API calls. Some README figures differ from the docs: the '0.21 ms p50' matches the docs table's p95 (p50 0.20 ms), and the docs tables from benchmarks/bench_latency.py show a repeated 500-string JSON array at 4.9K → 228 tokens but structured logs at 53% and 54%, against the README's '90%+ on repeated JSON arrays and log lines'. The 40–90% image-compression figure is stated without a benchmark.
Evidence: README.md:38, README.md:38-40, README.md:202-203
context-mode28.8% AI-signed
context-mode ships an offline benchmark suite (bench/run.mjs, 66 scenarios) with a committed results file, from which the README's 99.4% median byte saving recomputes (36 rows with 10 KB or more of raw output); these are byte counts on fixtures with tokens estimated as bytes / 4, not model runs. Live A/B runs in Claude Code with claude-sonnet-5-5 (scripts and committed aggregates; rerunning needs the `claude` CLI and paid model calls) found v2 cut the median tool_result bytes on large tasks by 84% (734.5 vs 4,694.5) with 12 of 12 large-task sessions passing in every arm, while session tokens were 5.7% worse than no plugin on small tasks and 19.6% worse on large ones; a later run put v2's session tokens 3.8% below no plugin on large tasks and 5.6% above on small ones. A paired cost study with committed run files and an analysis script found the open-source plugin cost more than Claude Code alone on six short tasks (v1 34.6% more, v2 2.2% more) and the hosted Context Mode Gateway's published rows 24.5% less, and the repo's own check of 13 numbers from the README's Try It section (commit a92cf6a7, measured on v1) found that 2 hold and 11 differ.
Evidence: README.md:29, README.md:415, bench/RESULTS.md:11
book-to-skill24.4% AI-signed
book-to-skill's headline, 24×–51× fewer tokens than placing the whole book in context to answer one question, comes from tools/discovery_tax.py run on three real books whose texts are not committed (119,264 / 175,253 / 256,287 context-dump tokens vs ~5,000 for book-to-skill, a ~4K core plus ~1K chapter that the tool takes as a design cap unless a generated skill is passed), and the 2.4×–15.6× discovery-loop figures are a model, as the tool's docstring states. Extraction counts, Docling vs pdftotext (164s vs 0.1s; 48 vs 0 tables), a ~$1-per-book generation cost and a one-chapter before/after table are published as results without a committed harness, and docs/performance.md lists 133 auto-detected chapters for Moby-Dick where docs/how-it-works.md lists none. The tools/evals harness for routing and answer accuracy has no live results yet, and docs/research/progressive-disclosure-evals.md summarizes an external study of a book-to-skill-based recipe in which a strong navigator on a single book reduced the accuracy benefit of the pre-built pack.
Evidence: README.md:107, docs/performance.md:42-55, docs/performance.md:14-25
ponytail24.1% AI-signed
Ponytail commits an agentic harness (benchmarks/agentic/run.py with judge, completeness and compare scripts) and results with per-task medians, 95% bootstrap intervals and sign tests: on 39 tasks with 5 runs per task and arm (585 sessions), Ponytail 5 reports -53% lines of code, -45% output tokens, -26% cost and -41% time versus no skill, and a test for 98% of risky logic (no skill: 68%). The same results report where it did not come out ahead: hidden checks are level (87/90 vs 86/90 for no skill, and all three arms passed 30 of 30 security runs), no-skill tests catch more injected bugs where a test exists (77% vs 69%), and a blind judge leaned toward the no-skill replies (82 : 106, p=0.09, not significant). An earlier single-shot benchmark found ponytail 26.2% and 38.7% more expensive than no skill on gpt-5.4-mini and gpt-5.5. Raw run workspaces are not committed, so checking the results requires a paid live rerun, and the /ponytail-review detection rates (100% vs 87% for no skill) appear in an image caption without data.
Evidence: README.md:42, benchmarks/README.md:13, benchmarks/results/2026-10-07-agentic.md:66
archify14.4% AI-signed
Archify publishes three committed 15-run matrices (3 models x 5 tasks) from its ordinary-model-floor benchmark, with per-run receipts and a runnable verifier and report harness. The first run recorded 10/15 first-pass usable; under the current verifier the original, post-fix and quality-first matrices each score 8/15, and the README states that the fixes do not demonstrate an overall uplift (lifecycle improved from 0/3 to 1/3 and architecture from 2/3 to 3/3, while workflow and data-flow regressed). Candidate generation uses an external runner and model providers that the repository does not ship, so the numbers can be re-verified from the committed candidates but not regenerated end to end from this checkout. A second harness, repair-rounds, has no published results.
Evidence: benchmarks/ordinary-model-floor/README.md:96-100, benchmarks/ordinary-model-floor/README.md:102-110, benchmarks/ordinary-model-floor/README.md:111
ECC7.9% AI-signed
ECC publishes GateGuard evaluations with runnable harnesses and committed aggregate results. Its pre-registered confirmatory run (run-intent.js, 96 trials) reports the gate's benefit as not confirmed (25/48 passed with the gate against 18/48 without, p = 0.210) at about 37% more turns and 40% more cost per trial; an earlier exploratory run gave 15/24 against 8/24 (p = 0.065). The replay and latency scripts report denials falling from 157 to 132 and must-deny bypasses from 8 to 0 on a 184-step corpus, and a fresh-process hook p50 of 23.68 ms against 11.58 ms on upstream/main; row-level results and transcripts of the hidden-intent runs are not committed. Outside GateGuard, the README's '~60% cost reduction' for defaulting to Sonnet is stated without data, and the mgrep '~2x fewer tokens' figure is attributed to Mixedbread.
Evidence: docs/gateguard/question-effectiveness.md:451, docs/gateguard/question-effectiveness.md:370, docs/gateguard/question-effectiveness.md:471
hermes-agent2.8% AI-signed
hermes-agent commits per-PR A/B evals under evals/ with runnable harnesses and scorecards that report token savings next to measured accuracy: lean+recovery compaction recall 68.3% vs 45.8% for the current policy at 3.3x fewer tokens (49K vs 162K; 96.7% uncompacted), core-tool deferral −7% to −23% tokens with grand accuracy 0.916 vs 0.923, and browser_exec −60% to −66% tokens with 18/18 runs succeeding on both arms. The scorecards also record results that went against the change, such as gpt-5.6-terra accuracy falling from 0.938 to 0.879 and the clarify tool being used in 7/18 runs instead of 18/18 under core-tool deferral. The runs need live paid models, the compaction transcripts are not committed, and per-run rows for the deferral, read-tool and browser suites are gitignored, so those results are summary tables; the session_search eval commits all 108 per-run rows (52/54 vs 49/54 passed). The toolperf figures (−21% turns, −29% tool calls on qwen3-coder-30b) are stated in a README next to the harness, without committed results.
Evidence: evals/compaction/results/SCORECARD-2026-08-15.md:18-21, evals/compaction/results/SCORECARD-2026-08-15.md:67, evals/core_tool_deferral/results/SUMMARY.md:1-20
OmniRoute1.7% AI-signed
OmniRoute's headline '15–95% tokens (~89% avg)' is arithmetic on third-party README figures: the docs combine RTK's 60-90% and Caveman's ~46% input savings as 1 − (1 − RTK) × (1 − Caveman_input), giving 78.4–94.6% and an 89.2% average. The per-mode figures (Lite ~15%, Standard ~30%, Aggressive ~50%, Ultra ~75%) and the 69 → 19 token (72%) example are stated without data. The harness-backed numbers are narrower: a committed CI floor of compressed tokens per task over a 5-case corpus (`npm run check:compression-budget`), which records no uncompressed counts and lists the same values (129 / 127 / 160) for lite, aggressive, rtk, session-dedup and headroom; a unit test that runs the single input behind the quoted 95.93% (300 identical error lines) and asserts savings above 90%; and a routing-event microbenchmark (`npm run bench:routing-events`, ~0.12 µs of event-dispatch overhead per request). Other harnesses (bench:compression, eval:compression, perf:lts-baseline) have no committed outputs.
Evidence: docs/compression/COMPRESSION_GUIDE.md:161, scripts/check/compression-budget-baseline.json:2-8, docs/architecture/ADAPTIVE_ROUTING.md:128
PageIndex0% AI-signed
The repo ships a real, re-runnable comparison notebook (63 vendor-written NVIDIA 10-K questions, LLM judge, saved outputs) showing PageIndex at 100.0% accuracy vs 42.9-92.1% for vector RAG configurations, though the vector answers and judge grades are not committed and the run depends on paid APIs. The headline README numbers (FinanceBench 98.7%, 97% token cut, financial-QA benchmark) point to external vendor repos/blogs, while the indexing cost/time and native-PDF cost ratios are plots with no committed harness.
Evidence: cookbook/pageindex-vs-vector-10k/pageindex-vs-vector-10k.ipynb:603, cookbook/pageindex-vs-vector-10k/pageindex-vs-embeddings-10k.ipynb:574, cookbook/pageindex-vs-vector-10k/pageindex-vs-vector-10k.ipynb:640
gentle-ai0.1% AI-signed
gentle-ai ships a deterministic, offline Go harness (bench/, its own go.mod, run in CI at .github/workflows/ci.yml:100-163) that measures the friction of its review lifecycle, such as human prompts, commands to completion and blocks, with no model calls and no timing, and commits one full run (bench/results.json: 64 journeys counted, 0 failed); a second results.json at the repo root records 0 counted and 1 failed journey. The README's 14-journey totals and its comparison with v2.1.2 (blocks 10 to 3 on the 5 comparable journeys) come from results-after.json and results-before.json, which are gitignored and not in the repo. The README's agent-performance table reports single (n=1) Gentle Shell runs with no data or harness in this repo; in it the data-migration run cost $0.555 against $0.403 before, and was accepted on the first round. The spec row and the comparison with Codex CLI link to gentle-shell PRs and an issue (#1718, #1723, #1794, #1713); the comparison reports that the large task took about three times longer (17.5–17.9 min vs 6.1 min).
Evidence: bench/results.json:12558, bench/README.md:409-431, bench/README.md:439-446
All 52 repositories
| Repository | Stars | Created | AI evidence | AI-signed % | AI-signed % (last 180d) | Fix % | Churn | Survival % | Published results |
|---|---|---|---|---|---|---|---|---|---|
| superpowers | 314,290 | 2025-10-09 | AI-signed | 13.3 | 6.1 | 24.5 | 3.86 | 41.2 | Results, no harness |
| ECC | 299,342 | 2026-01-18 | AI-signed | 8.1 | 7.9 | 43.7 | 1.57 | 77.8 | Harness + committed results |
| deepseek-harness | 276,014 | 2026-08-13 | Agent config, <5% signed | 0.0 | 0.0 | 38.4 | 2.43 | 57.9 | Harness + committed results |
| hermes-agent | 261,582 | 2025-07-22 | Agent config, <5% signed | 2.9 | 2.8 | 53.5 | 2.73 | 53.8 | Harness + committed results |
| firecrawl | 199,609 | 2024-04-15 | AI-signed | 13.1 | 42.1 | 32.3 | 2.16 | 63.2 | Results, no harness |
| AutoGPT | 187,937 | 2023-03-16 | AI-signed | 20.4 | 73.6 | 36.7 | 2.55 | 56.2 | Harness + committed results |
| ponytail | 186,540 | 2026-06-12 | AI-signed | 24.1 | 24.1 | 41.8 | 1.66 | 75.2 | Harness + committed results |
| ollama | 186,319 | 2023-06-26 | Agent config, <5% signed | 0.1 | 0.2 | 15.1 | 4.45 | 36.7 | Harness, no committed results |
| transformers | 168,646 | 2018-10-29 | AI-signed | 1.0 | 10.0 | 31.7 | 3.38 | 45.7 | Harness + committed results |
| dify | 160,649 | 2023-04-12 | AI-signed | 5.3 | 6.7 | 42.6 | 3.55 | 43.9 | Harness, no committed results |
| open-webui | 158,080 | 2023-10-06 | Neither signal (unknown) | 1.3 | 2.6 | 22.2 | 3.37 | 45.8 | Claim without data |
| langchain | 149,190 | 2022-10-17 | Agent config, <5% signed | 0.3 | 2.1 | 21.7 | 5.63 | 30.1 | Harness, no committed results |
| graphify | 144,135 | 2026-04-03 | AI-signed | 50.4 | 52.7 | 56.6 | 1.4 | 83.2 | Harness + committed results |
| browser-use | 120,289 | 2024-10-31 | AI-signed | 4.9 | 5.6 | 22.2 | 5.56 | 30.5 | Results, no harness |
| orca | 118,822 | 2026-03-17 | AI-signed | 14.7 | 13.2 | 46.6 | 3.63 | 43.2 | Harness + committed results |
| caveman | 118,218 | 2026-04-04 | AI-signed | 78.7 | 82.3 | 55.5 | 1.8 | 71.5 | Harness + committed results |
| claude-mem | 114,633 | 2025-08-31 | AI-signed | 61.7 | 66.9 | 46.6 | 3.38 | 47.0 | Harness + committed results |
| archify | 94,563 | 2026-04-15 | AI-signed | 14.0 | 14.4 | 35.8 | 1.43 | 81.5 | Harness + committed results |
| ragflow | 93,749 | 2023-12-12 | Agent config, <5% signed | 2.4 | 4.3 | 49.3 | 3.03 | 49.6 | Results, no harness |
| rtk | 85,862 | 2026-01-22 | AI-signed | 34.8 | 35.2 | 60.1 | 1.79 | 71.8 | Harness, no committed results |
| OmniRoute | 85,531 | 2026-02-13 | Agent config, <5% signed | 3.0 | 1.7 | 61.2 | 2.3 | 51.9 | Harness + committed results |
| lobehub | 84,366 | 2023-05-21 | AI-signed | 22.0 | 40.7 | 34.9 | 1.87 | 69.7 | Harness, no committed results |
| headroom | 79,105 | 2026-01-07 | AI-signed | 31.0 | 32.3 | 63.2 | 1.58 | 77.4 | Harness + committed results |
| OpenSpec | 74,248 | 2025-08-05 | AI-signed | 30.6 | 63.9 | 42.9 | 1.66 | 75.1 | Claim without data |
| anything-llm | 68,053 | 2023-06-04 | Neither signal (unknown) | 0.4 | 0.5 | 25.3 | 2.0 | 66.7 | Claim without data |
| humanizer | 66,094 | 2026-01-18 | AI-signed | 32.4 | 35.8 | 25.4 | 6.07 | 28.3 | Third-party only |
| get-shit-done | 64,157 | 2025-12-14 | AI-signed | 50.0 | 50.0 | 43.3 | 1.82 | 70.5 | Results, no harness |
| academic-research-skills | 55,793 | 2026-02-26 | AI-signed | 75.8 | 70.4 | 27.9 | 1.16 | 92.7 | Harness + committed results |
| BMAD-METHOD | 55,118 | 2025-04-13 | AI-signed | 14.7 | 8.5 | 31.7 | 35.84 | 5.4 | Harness, no committed results |
| herdr | 49,219 | 2026-03-27 | Agent config, <5% signed | 0.3 | 0.3 | 55.1 | 1.84 | 70.4 | Harness, no committed results |
| agents | 42,294 | 2025-07-24 | AI-signed | 11.9 | 20.5 | 29.8 | 2.26 | 58.5 | No quantitative results |
| PageIndex | 40,181 | 2025-04-01 | Neither signal (unknown) | 0.6 | 0.0 | 25.0 | 1.39 | 83.6 | Harness + committed results |
| QwenPaw | 37,473 | 2026-02-24 | Neither signal (unknown) | 0.3 | 0.2 | 43.3 | 1.58 | 77.5 | Results, no harness |
| book-to-skill | 37,242 | 2026-05-01 | AI-signed | 27.2 | 24.4 | 42.7 | 1.28 | 87.5 | Harness + committed results |
| context-mode | 33,662 | 2026-02-23 | AI-signed | 36.3 | 28.8 | 31.6 | 3.5 | 44.2 | Harness + committed results |
| go-zero | 33,420 | 2020-08-07 | AI-signed | 1.3 | 23.1 | 26.7 | 2.29 | 60.7 | Third-party only |
| cmux | 28,745 | 2026-01-28 | AI-signed | 24.5 | 27.2 | 35.0 | 1.65 | 74.4 | Harness + committed results |
| happy | 24,536 | 2025-07-18 | AI-signed | 13.9 | 5.2 | 24.2 | 2.27 | 61.2 | Harness + committed results |
| Paradev | 16,090 | 2025-10-21 | AI-signed | 5.1 | 4.1 | 40.6 | 2.34 | 59.8 | Harness + committed results |
| agent-orchestrator | 14,256 | 2026-02-13 | AI-signed | 40.4 | 34.9 | 58.6 | 1.92 | 68.5 | Harness + committed results |
| elephant-goldfish-model | 14,254 | 2026-09-13 | Agent config, <5% signed | 0.7 | 0.7 | 26.6 | 1.46 | 81.4 | No quantitative results |
| holaOS | 11,434 | 2026-03-22 | AI-signed | 85.2 | 85.2 | 42.6 | 1.04 | 98.2 | Results, no harness |
| gsd-core | 11,334 | 2026-05-22 | AI-signed | 48.5 | 48.0 | 49.3 | 1.82 | 70.6 | Harness + committed results |
| munder-difflin | 9,425 | 2026-05-31 | AI-signed | 72.6 | 72.6 | 34.5 | 4.37 | 37.2 | Harness, no committed results |
| Orkas | 9,098 | 2026-04-29 | AI-signed | 5.2 | 2.2 | 26.7 | 1.18 | 91.5 | Results, no harness |
| openrig | 8,737 | 2026-04-01 | AI-signed | 26.7 | 19.5 | 49.1 | 1.34 | 85.4 | Harness, no committed results |
| gentle-ai | 8,586 | 2026-02-27 | Agent config, <5% signed | 0.3 | 0.1 | 55.1 | 2.34 | 59.8 | Harness + committed results |
| yao | 8,280 | 2021-09-06 | AI-signed | 3.7 | 33.0 | 11.6 | 2.51 | 57.0 | Harness + committed results |
| failproofai | 8,230 | 2026-04-06 | AI-signed | 83.7 | 79.3 | 32.6 | 2.32 | 60.3 | Harness, no committed results |
| thinkrail | 7,921 | 2026-07-08 | AI-signed | 6.1 | 6.1 | 31.0 | 1.57 | 77.9 | Harness, no committed results |
| gsd-2 | 7,750 | 2026-03-11 | AI-signed | 25.7 | 25.7 | 64.6 | 1.79 | 71.1 | Harness, no committed results |
| conductor | 3,777 | 2025-12-17 | Neither signal (unknown) | 0.0 | 0.0 | 33.3 | 1.92 | 68.6 | No quantitative results |
Fix %, churn ratio and survival are measured at HEAD. Click a column to sort. The same table is available as CSV; the per-repo explorer with commit timelines is in the academic version.
Method and limits
How it was measured
- The 52 repos. The top 10 by stars in each of 7 GitHub topics (
agent-harness,ai-agents,coding-agents,context-engineering,llm,prompt-engineering,spec-driven-development) as listed on 2026-11-08, minus reading material and collections (books, courses, guides, prompt collections, awesome-lists), plus 4 repos the work started from (OmniRoute, context-mode, rtk, superpowers). 39 were created in 2025 or later, 11 before 2024 (AutoGPT, anything-llm, dify, go-zero, langchain, lobehub, ollama, open-webui, ragflow, transformers, yao). Median stars 65,126 (middle half: 15,632–126,250), fetched 2026-11-08. - The baseline. denoland/deno, excalidraw/excalidraw, fastapi/fastapi, junegunn/fzf, gin-gonic/gin, gradio-app/gradio, hoppscotch/hoppscotch, jesseduffield/lazygit, n8n-io/n8n, microsoft/playwright, BurntSushi/ripgrep, huggingface/transformers. Both groups are chosen by today's fame. Snapshots that predate Copilot (go-zero, transformers, yao at day 180) are left out of the age comparison, so no repo counts on both sides.
- Commits. Full history of the default branch. 38,229 merge commits and 10,346 maintenance-bot commits (dependabot, release bots) are left out of every percentage. get-shit-done and gsd-core share 2,554 commits of history (up to 2026-05-22), which count in both.
- AI signals. Signed: an AI co-author trailer, an AI agent as author or committer, or a tool's “Generated with” footer. Configured: the repo commits agent instructions (
CLAUDE.md,AGENTS.md,.cursor/,.github/copilot-instructions.md…) or runs a coding agent in CI. - Fix % is the share of commits whose subject contains fix, bug, hotfix or regression. Churn ratio is lines added plus deleted over all history, divided by lines at HEAD. Survival % is lines at HEAD divided by lines ever added (source code and Markdown; vendored, generated, translated, lock and changelog files excluded).
- Published results are each repo's own claims about its performance, located by file and line and classified by evidence type.
What this does not show
- No causation. Famous projects were selected by today's stars, so they are survivors. Projects that used AI the same way and failed are not in the sample.
- AI-tooling projects, not famous projects in general. People who build tools for coding agents are the likeliest to use them.
- A snapshot of moving lists. Topic rankings change daily; this is the top 10 of each topic on 2026-11-08.
- AI share is a lower bound. Signatures measure attribution, not authorship, and Claude Code is over-represented because it signs by default.
- Agent configuration can be product, not process. In AI-tooling repos a root
CLAUDE.md,.claude/or.cursor/can be a template the product ships rather than instructions for the project's own development. - Signed is not unreviewed. A signature says nothing about how carefully a person reviewed the commit.
- AI use and recency go together. Almost all heavy AI use is in projects started in 2025 or later, so differences from older projects cannot be attributed to AI rather than to when they started.
- Fix % is keyword-based and sensitive to commit-message conventions; it is not a defect rate. Churn mixes everything that is committed, including vendored or generated files that escape the exclusion rules.
- Published results are self-reported, located and classified but not reproduced.
- Small baseline. 12 pre-Copilot projects, chosen by fame, language and clone size.