Agentic Very Much: Coding Agent Adoption Has Doubled in New GitHub Projects — What It Means for Codex CLI Teams

Updated

codex-clicoding-agentsadoptiongithubAGENTS.mdcommit-attributionteam-configurationresearch

Agentic Very Much: Coding Agent Adoption Has Doubled in New GitHub Projects — What It Means for Codex CLI Teams

A follow-up study from Robbes, Matricon, Degueule, Hora, and Zacchiroli — published on arXiv on 5 June 2026 — finds that coding agent adoption in newly created GitHub projects is more than twice as high as the already significant levels measured in their predecessor study six months earlier 1. The proportion of AI-assisted commits is also substantially higher, and the authors acknowledge strong signals that their detection heuristics still miss a significant fraction of agent activity 1. For teams running Codex CLI, this is not merely a curiosity statistic: it reshapes assumptions about commit provenance, review load, and the configuration posture every repository should adopt.

The Numbers: From One-in-Five to Nearly One-in-Two

The predecessor study, “Agentic Much?” (arXiv:2601.18341, January 2026), analysed 129,134 GitHub projects with at least 10 stars, 5,000 lines of code, and 100 commits 2. It reported a conservative file-level adoption rate of 15.85% and an upper estimate of 22.60% 2. Among the 10,191 projects showing file-level adoption signals, commit-level adoption broke down as follows 2:

  • Experimental (<1% of commits): 7.3%
  • Limited (1–5%): 16.1%
  • Consistent (5–20%): 17.6%
  • Pervasive (>20%): 17.8%

The “Agentic Very Much!” follow-up analyses projects created after that dataset closed and finds adoption more than doubled 1. Crucially, the intensity is also higher: a larger share of commits carry agent attribution in the newer cohort 1. Given the predecessor’s acknowledged undercount — 41.2% of file-level adopters showed zero commit-level traces despite having configuration files — the true adoption intensity is almost certainly higher still 2.

Contextual industry data aligns: 84% of developers now use or plan to use AI coding tools, with 51% reporting daily usage, yet only 29% trust the output 3. GitHub Copilot alone serves 140,000 organisations 4. The broader coding agent category — which includes Codex CLI, Claude Code, Cursor, and dozens of others — grew commit volume by 1,400% in 2026 5.

The Detection Problem: What You Cannot See

The studies used four heuristic categories to detect agent use 2:

graph TD
    A[Detection Heuristics] --> B[File-based<br/>86 patterns]
    A --> C[Commit-based<br/>20 patterns]
    A --> D[Branch-based<br/>20 patterns]
    A --> E[Label-based<br/>44 patterns]

    B --> B1[CLAUDE.md]
    B --> B2[AGENTS.md]
    B --> B3[.cursorrules]
    B --> B4[copilot-instructions.md]

    C --> C1[Co-authored-by trailers]
    C --> C2[Author metadata]
    C --> C3[Commit message patterns]

    D --> D1[Branch naming conventions]

    E --> E1[PR/issue labels]

The file-based category is the most reliable: if a repository contains AGENTS.md, CLAUDE.md, or .cursorrules, it is almost certainly using agents. But commit-based detection depends entirely on the agent leaving a trail — and several tools, including Codex CLI, allow attribution to be silenced.

The commit_attribution configuration key in Codex CLI’s config.toml defaults to "Co-authored-by: Codex <noreply@openai.com>" 6. Setting it to an empty string removes the trailer entirely:

# config.toml — RECOMMENDED: keep attribution visible
[git]
commit_attribution = "Co-authored-by: Codex <noreply@openai.com>"

# ANTI-PATTERN: silencing attribution
# commit_attribution = ""

When attribution is silenced, the only remaining detection signal is behavioural analysis. MSR 2026 research by Ghaleb analysed 33,580 pull requests from five agents and achieved a 97.2% F1-score in multi-class identification using 41 features — the dominant discriminator being commit message style rather than code quality or diff size 7. But without explicit trailers, passive detection accuracy drops to roughly 60% 7.

Why Transparent Attribution Matters Now

The doubling of adoption means that in any team of reasonable size, agent-assisted commits are no longer exceptional — they are the emerging norm. Three consequences follow:

1. Audit Compliance

Regulated industries — finance, healthcare, government contracting — increasingly require logging where AI tooling was used in software delivery 6. A Co-authored-by trailer in every relevant commit provides a machine-readable audit trail with zero additional tooling. For Codex CLI, the default configuration already handles this; the risk lies in teams deliberately or accidentally disabling it.

2. Review Calibration

Agent-assisted commits tend to be larger than human-only commits and disproportionately contain features and bug fixes 2. Reviewers who do not know a commit is agent-generated may apply incorrect heuristics — either over-trusting (assuming a senior developer wrote it) or under-scrutinising (rubber-stamping large diffs). Visible attribution enables reviewers to switch to agent-appropriate review strategies.

3. Adoption Measurement

You cannot manage what you cannot measure. Teams adopting Codex CLI should be able to answer: “What percentage of our commits are agent-assisted this sprint?” Without attribution, this question requires expensive after-the-fact analysis.

Configuring Codex CLI for the Post-Adoption-Doubling World

Enforce Attribution at the Repository Level

Use an AGENTS.md at the repository root to mandate attribution and prevent silencing:

<!-- AGENTS.md -->
## Git Commit Policy

- All agent-generated commits MUST include a Co-authored-by trailer
- Do NOT set commit_attribution to empty string
- Use the default trailer or the team-specific format below:
  Co-authored-by: Codex <noreply@openai.com>

Pair this with a git commit-msg hook that rejects commits matching agent behavioural patterns but lacking a Co-authored-by trailer:

#!/usr/bin/env bash
# .git/hooks/commit-msg — reject probable agent commits without attribution
MSG=$(cat "$1")
if echo "$MSG" | grep -qiE '(refactor|implement|add support for|fix issue)' && \
   ! echo "$MSG" | grep -qi 'Co-authored-by'; then
  echo "ERROR: Commit message matches agent patterns but has no Co-authored-by trailer."
  echo "If this is a human-authored commit, add --no-verify to override."
  exit 1
fi

Track Adoption Intensity with /usage and ccusage

Codex CLI v0.140.0 shipped the /usage command for daily, weekly, and cumulative token activity 8. Combine this with external trackers like ccusage, which reads Codex session JSONL files under ~/.codex, to build a team-level adoption dashboard 9:

flowchart LR
    A[Codex CLI Sessions] -->|JSONL logs| B[ccusage]
    C[Git Commits] -->|Co-authored-by| D[git log --grep]
    B --> E[Team Dashboard]
    D --> E
    E --> F[Adoption Intensity<br/>Metric: % agent commits]
    E --> G[Token Spend<br/>Metric: $/developer/week]

Use AGENTS.md as Both Configuration and Signal

AGENTS.md now serves dual purposes. It is a configuration file that Codex CLI reads before starting work, and it is a detection signal that researchers and internal auditors use to identify agent-enabled repositories 2. As of mid-2026, AGENTS.md has been adopted by over 60,000 open-source repositories and is stewarded by the Linux Foundation’s Agentic AI Foundation alongside MCP 10.

For teams, this means AGENTS.md should be treated with the same discipline as .editorconfig or CODEOWNERS — committed, reviewed, and enforced in CI:

# config.toml — named profile for CI verification
[profiles.ci-check]
model = "o4-mini"
approval_policy = "unless-allow-listed"

Configure Named Profiles for Adoption Phases

The predecessor study’s adoption intensity categories — experimental, limited, consistent, pervasive — map naturally to Codex CLI named profiles 2:

Adoption Phase Profile approval_policy Typical Use
Experimental cautious suggest Test generation, documentation
Limited standard auto-edit Bug fixes, small features
Consistent trusted full-auto Established patterns, CI-gated
Pervasive autonomous full-auto + hooks Full SDLC with PostToolUse gates 
# config.toml — progressive adoption profiles
[profiles.cautious]
model = "o4-mini"
approval_policy = "suggest"

[profiles.trusted]
model = "o3"
approval_policy = "full-auto"

The Underdetection Imperative

The most striking finding in the “Agentic Very Much!” study is not the doubling itself but the authors’ candid acknowledgement that their heuristics systematically undercount agent activity 1. If the conservative estimate in October 2025 was 15.85%, the true rate was likely 20–25%. If the new cohort shows double that, the actual adoption in newly created projects may exceed 50%.

This has a direct operational implication: assume agent-assisted code is present in every repository you touch. Configure your review workflows, CI gates, and compliance tooling accordingly — not as a response to confirmed agent use, but as a default posture.

The era where agent-assisted development was optional, experimental, or niche is over. The data says it doubled. The undercount says it is probably worse. Configure accordingly.

Citations

  1. Robbes, R., Matricon, T., Degueule, T., Hora, A., & Zacchiroli, S. (2026). “Agentic Very Much! Adoption of Coding Agent in New GitHub Projects.” arXiv:2606.07448. https://arxiv.org/abs/2606.07448 ↩ ↩2 ↩3 ↩4 ↩5

  2. Robbes, R., Matricon, T., Degueule, T., Hora, A., & Zacchiroli, S. (2026). “Agentic Much? Adoption of Coding Agents on GitHub.” arXiv:2601.18341. https://arxiv.org/abs/2601.18341 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  3. Uvik Software. (2026). “AI Coding Assistant Stats 2026: 84% Adoption, 29% Trust.” https://uvik.net/blog/ai-coding-assistant-statistics/ ↩

  4. Panto AI. (2026). “GitHub Copilot Statistics 2026 — Users, Revenue & Adoption.” https://www.getpanto.ai/blog/github-copilot-statistics ↩

  5. Digital Applied. (2026). “AI Coding Adoption 2026: 50 Statistics From 7 Surveys.” https://www.digitalapplied.com/blog/ai-coding-adoption-statistics-2026-50-data-points ↩

  6. Codex Knowledge Base. (2026). “Codex CLI Commit Attribution: Tagging Agent Work with commit_attribution.” https://codex.danielvaughan.com/2026/03/28/codex-cli-commit-attribution/ ↩ ↩2

  7. Codex Knowledge Base. (2026). “Agent Fingerprints in Pull Requests: What MSR 2026 Research Reveals and How to Configure Codex CLI for Professional Git Hygiene.” https://codex.danielvaughan.com/2026/04/30/agent-fingerprints-pull-requests-codex-cli-git-hygiene/ ↩ ↩2

  8. OpenAI. (2026). “Codex CLI Changelog — v0.140.0.” https://github.com/openai/codex/releases ↩

  9. ccusage. (2026). “Coding (Agent) CLI Usage Analysis.” https://ccusage.com/guide/codex/ ↩

  10. Linux Foundation. (2025). “Linux Foundation Announces the Formation of the Agentic AI Foundation (AAIF).” https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation ↩