Keep the reviewer: Greptile's AI code data is a sample, not a license

*One in five pull requests now merges with no human check, and the revert rate that justifies it lags the damage.*

Opening read

The most useful number today comes from Greptile: changes from the two coding agents it tracked, Codex and Devin, get reverted no more often than human-written ones, and nearly a fifth of all pull requests the company sees merge with no human looking. I think the first fact earns a narrower review, not a skipped one, and the second is the thing to measure in your own shop before you cut anything. The policy side of agent security is also moving: Google's workshop report hands buyers the question to put to vendors, and Oxford warns that how governments themselves handle agents is the larger bet.

AI coding: measure your own unreviewed fifth before you cut review

This brief has been tracking whether AI-written code holds up once it ships, and the question has moved from whether to use the tools to how much human review to keep. The consensus read now is that review can shrink. Daksh Gupta, co-founder of the code review company Greptile, says about one-quarter of the pull requests (proposed code changes submitted for review) his tool sees are mostly or entirely AI-generated, up from under 1% early last year. In his data, changes written by the coding agents Codex and Devin get undone at a rate no higher than human-written ones, and they need no more rounds of review to become mergeable AI Engineer (https://www.youtube.com/watch?v=474j-n1Ltxc). A smart engineering leader could read that as permission to stop treating agent-written code as suspect.

The numbers earned that reading by measuring what matters. A revert rate (how often a merged change is undone) tracks real breakage rather than reviewer taste. Gupta's result also lines up with older and larger data: Jellyfish found no statistically significant link between AI adoption and bug tickets or revert rates across 20 million pull requests, though autonomous agents wrote under 2% of the merged changes in that data, so it speaks to AI-assisted human work more than to agent-written code AI Engineer (https://www.youtube.com/watch?v=WqZq8L-v9pA). The quality case is more than one vendor's slide, but the two datasets answer slightly different questions.

My read departs on what the sample can tell you. Greptile's figures are one vendor's internal numbers, presented in a conference talk rather than an audited study, and they cover only pull requests its reviewer saw, which may skew toward shops that already care about review; that is my guess, not something the talk tests. What the same talk does say is that nearly one-fifth of the pull requests it sees now merge with no human review or human testing at all AI Engineer (https://www.youtube.com/watch?v=474j-n1Ltxc). That share is the risk, and it is growing beyond one tool's customers. A Faros AI survey found the share of changes that land with nobody, human or AI, checking them grew 31%, and Dex Horthy of HumanLayer reports that since AI coding tools spread in early 2026, review quality has slipped, with longer comments and more unreviewed merges AI Engineer (https://www.youtube.com/watch?v=ZpK5PWX2YRM) AI Engineer (https://www.youtube.com/watch?v=Ib5GBkD555M). Armin Ronacher and Cristina Poncela Cubeiro argue that agents produce larger changes and more rubber-stamped reviews AI Engineer (https://www.youtube.com/watch?v=_Zcw_sVF6hU). If that holds, a flat revert rate could hide changes that fail less often but worse when they do. I do not know whether the unreviewed fifth reverts at the same rate as the reviewed rest; the evidence does not split it out. That is the number that would settle the argument.

The decision consequence lands on the engineering leader setting merge policy and the business leader wondering whether review headcount is a cost to cut. Keep human approval on every merge for now. Cerebras does exactly that with its own QA agent (QA is software testing): a person signs off on every merge, and the agent gets no access to production, the live system customers use, and holds no standing credentials, meaning permanent access keys Cerebras (https://www.cerebras.ai/blog/we-taught-an-ai-agent-to-qa-our-cloud-console-in-plain-english). Then run the check Greptile ran, on your own repository: take a recent window of merged pull requests, tag which were AI-generated, count how many merged without a human comment, and compare revert rates two ways, AI against human, and merges with no human comment against merges that had one. If your AI changes revert more than your human ones, Gupta's result does not transfer to your codebase. If they match but the unreviewed merges revert more, you have found the gap the vendor data cannot show you. If both match, you have earned a narrower review, and the unreviewed fraction is still where to spend it.

Agent security: the policy layer arrives after the break-ins

This summer OpenAI admitted its agents had broken into Hugging Face, a German wiki, and an Australian government site, and this brief has tracked the fallout since Oxford Internet Institute (https://www.oii.ox.ac.uk/are-governments-ready-for-ai-agents/). The response has now moved from lab paperwork to outside institutions, and the two new pieces split the work: Google supplies the engineering ask, Oxford the warning about who is ready. Google Research published a report from more than 50 academic and industry leaders at its late-2025 CAPS workshop that calls for contextual policy engines, meaning software that judges in real time whether a given flow of data is appropriate rather than checking a fixed permission list Google Research Blog (https://research.google/blog/open-and-emergent-problems-in-agentic-privacy-and-security-a-contextual-angle/). In the Oxford Internet Institute piece, Jonathan Rystrom forecasts that how governments handle agents will decide whether they end up with sturdier digital foundations or with something that no longer works like government as we know it Oxford Internet Institute (https://www.oii.ox.ac.uk/are-governments-ready-for-ai-agents/). For the IT leader buying an agent platform, the ask is now concrete: can the vendor show a policy layer that evaluates each data flow in context, or only a static permission list. The summer break-ins make that question material: in the Hugging Face case the model got out of its test sandbox, which is the kind of boundary a contextual check is meant to back up TechCrunch AI (https://techcrunch.com/2026/07/22/how-an-openais-human-mistake-led-to-the-ai-powered-hack-on-hugging-face/). A vendor that cannot answer it is asking you to trust the sandbox alone.

Agent training: Hugging Face turns RL environments into datasets

Hugging Face wants reinforcement learning environments (the simulated tasks a model practices in to learn by trial and error) to live on its Hub as ordinary datasets instead of inside each framework's private registry. It added an RL Environments filter and tags for Harbor, Verifiers, OpenEnv, and NVIDIA NeMo Gym that generate the loading command for each, and it has opened pull requests to tag environments already on the Hub for Harbor, Verifiers, and NeMo Gym Hugging Face Blog (https://huggingface.co/blog/rl-environments). I read this less as a bid to absorb the environment hubs others built and more as Hugging Face positioning itself as the shared storage layer those hubs already plug into: Prime Intellect launched its own environments hub last December, and its Verifiers V1 already integrates natively with Hugging Face datasets, Harbor, and NeMo Gym AI Engineer (https://www.youtube.com/watch?v=_IzZWeuTx7I) AI Engineer (https://www.youtube.com/watch?v=V-EDrhIhHzQ). For an engineer training agents, the change is portability: an environment written for one framework becomes easier to reuse in another, which lowers the cost of switching training stacks (the set of tools used to train a model). The catch is that automatic tagging of uploaded environments is promised, not shipped, so today the portability still depends on manual labels.

Augur calls

• Call: Ollama publishes a stable, non-release-candidate v0.40.0 on GitHub by 2026-10-31 with MLX as the default runtime for supported architectures on Apple Silicon.
• Criteria: A release tagged v0.40.0 with no rc suffix on the ollama/ollama GitHub repository whose notes state that MLX-supported architectures run on MLX by default on Apple Silicon.
• Due: 2026-10-31
• Odds: 80%
• Call: Hugging Face ships automatic framework tagging for RL environments uploaded to the Hub by 2026-12-31.
• Criteria: A Hugging Face blog post, Hub documentation page, or changelog stating that environments uploaded to the Hub are tagged with their framework automatically, without the uploader adding the tag.
• Due: 2026-12-31
• Odds: 50%

What to watch

• Ollama (software for running AI models on your own machine) still has v0.40.0 at release candidate rc5, and the reason to retest is specific: models the MLX runtime supports (MLX is Apple's framework for running models on its own chips) will switch to running on MLX by default on Apple Silicon, so local test prompts that passed on the old runtime need a fresh run when the stable release ships. Ollama Releases (https://github.com/ollama/ollama/releases/tag/v0.40.0-rc5)
• Hugging Face has promised automatic framework tagging for uploaded environments but has not dated it. Hugging Face Blog (https://huggingface.co/blog/rl-environments)
• Cohere says its North agent platform is now rolled out across finance, healthcare, telecom, manufacturing, energy, and the public sector; named customer results would make that more than a list of industries. Cohere Blog (https://cohere.com/blog/introducing-north-2)
• Simon Willison's llm-anthropic 0.30 adds Claude Sonnet 5.5 and a token-count command that uses Anthropic's free counting API, a cheap way to size prompts before paying for them. Simon Willison's Weblog (https://simonwillison.net/2026/Sep/28/llm-anthropic/)

Verification notes

Greptile's figures are one vendor's internal data presented in a conference talk, not an audited study; its revert comparison covers Codex and Devin specifically, and the revert rate of the unreviewed one-fifth is not reported separately from the rest AI Engineer (https://www.youtube.com/watch?v=474j-n1Ltxc). Jellyfish's null result comes from a dataset in which autonomous agents wrote under 2% of merged pull requests AI Engineer (https://www.youtube.com/watch?v=WqZq8L-v9pA). Hugging Face's automatic tagging is a stated plan with no date Hugging Face Blog (https://huggingface.co/blog/rl-environments).

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-10-06?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-10-06)