Engineer John Wang, CTO of Assembled, published an analysis of the open openai/codex repository as a verifiable 'ground truth' of agentic development. According to git history data, the Rust part of the project grew from 98 commits from 6 authors in May 2025 to more than 1000 commits from 135 authors in the first 25 days of August 2026 with a team of 137 people. The 8–10x growth, according to the analysis, is explained by a combination of massive use of coding agents, aggressive hiring, and guardrails codified in AGENTS.md and automatic lint and CI checks. All described practices are open and available for free copying.

image
image
image

What happened

John Wang studied the open openai/codex repository by its git history and published an analysis on August 27, 2026. The data shows a change in scale: in May 2025, the Rust part received 98 commits from 6 authors, of which 150 of the first 169 Rust commits were written by one engineer, Michael Bolin; in the first 25 days of August 2026, the repository received more than 1000 commits from 135 authors with a team of 137 people. The project has a 322-line AGENTS.md file that codifies observed agent bypass behavior: it is forbidden to modify the CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR checks and to create meaningless tests for the sake of formally passing CI. Any change to agent logic requires an integration test through the TestCodexBuilder harness. The rule not to use bool and Option parameters without a comment /*parameter_name*/ is supported by a custom lint that appeared in March 2026; there are 38 lint rules in the repository in total.

Context

openai/codex is OpenAI's open Codex CLI in Rust and TypeScript with approximately 121k stars on GitHub, making it one of the most notable open projects in the coding agents category. The value of the analysis is that it is based on open data, not conference statements: the scale of contributions, AGENTS.md rules, and automatic checks can be re-verified by git history, not taken on faith. The OpenAI approach to fixing changes is reflected in the TUI migration process: the tui_app_server feature flag was created on March 16, 2026, enabled by default after 10 days, after which the old implementation was removed, and CI prohibited importing codex-core. The migration is maintained not by team memory, but by automation. The test infrastructure is built just as thoroughly: tests take up about 615k lines, about 40% of the codebase, including a stable Responses API of about 7k lines that runs a real agent cycle against fake model responses.

Why this matters for the industry

This is a rare case where agentic development practices are verifiable by open data, not conference presentations. OpenAI converts recurring code review comments into AGENTS.md rules, and then into deterministic lint and CI checks, delegating control to automation instead of model or team memory. It was precisely this loop that allowed 135 authors and agents to simultaneously edit about 28 Rust crates on a median active day. For builders, this is a signal that the market opportunity now is not in another coding agent, but in infrastructure: lint rule libraries for agent repositories, harnesses with fake API stubs as a separate product, tools for validating agent PRs for meaningless changes. If the 'code review comment, AGENTS.md rule, lint/CI' combination gets community validation, it will likely begin to be copied as a de facto standard, and agentic development governance risks becoming mandatory engineering infrastructure at the CI/CD level.

Why this matters for users

All patterns are immediately and freely transferable, without new models and budgets, and can be implemented in days in existing repositories. A team using coding agents should create an AGENTS.md with explicit prohibitions on typical bypasses, including modifying sandbox checks and tests for the sake of tests, and include a lint for ambiguous literal arguments in mandatory checks. The test run mode is fully transferable: during development, run only the tests of the affected crate, and run the full Cargo set across five platform-architecture combinations after merging. The Responses API stub from the repository is a ready-made example of how to check the agent cycle without a live model call. The openai/codex repository and Wang's analysis itself serve as a checklist for teams implementing coding agents and wanting to evaluate them honestly.

What is not yet known / limitations

The main caveat concerns causality: the compared windows are methodologically unequal, the full month of May 2025 versus the first 25 days of August 2026, and the three explanatory factors, massive use of coding agents, hiring, and guardrails, remain an observational correlation without a counterfactual. The data shows that the growth coincided with these changes, but does not prove that they caused it. All figures come from Wang's analysis and are open to re-verification by the repository's git history, but their interpretation belongs to the author of the analysis, not OpenAI.

Sources

Author

Look at AI, editorial team