The agents wrote 200 services. Nobody reviewed a line.

The factory arrives

AWS has released a local testing environment for more than 200 of its services, and the striking part is not the product. It is how it was built. Qodo reports the AWS Emulator was written entirely by autonomous agents without human code review. Instead of line-by-line review, the project relies on a robust automated merge gate that includes security agents performing adversarial reviews. The engineer's job, in this workflow, is to define the rules and build the merge gates. The limits are stated plainly too: the emulator does not evaluate IAM policies the way production AWS does, and AWS advises against using it for security testing. The agents built the system. The humans built the test that decides whether the system ships. m:…3362

The models have earned that trust at the level of mechanics. Snorkel reports its Fable 5.2 model achieved a 71 percent pass-at-one score on a set of 26 frontier coding tasks, up from 53 percent for Fable 5.1. Mid-run terminal failures observed in the earlier version are nearly absent, and the remaining failures are concentrated in planning and reasoning — stable misreadings of task requirements — rather than crashes and malformed output. The model recorded 302 instances of self-verification, with 71 percent rated as strong. The agents have largely stopped failing at typing. They fail, now, at understanding what was asked. m:…6119

Verification is the bottleneck

Qodo's argument this week is that most organisations calling their setup a software factory are getting the definition wrong. A factory, it says, is a delivery system where AI agents plan, build, review and release software within defined boundaries, checks and evidence — not a workflow with agents in it. The bottleneck in such a system is verification, not code generation, and teams that add agents to every stage without improving review capacity often end up with longer queues. The stakes are quantified by Gartner, cited by Qodo: by 2028, half of enterprises adopting autonomous coding agents without real-time guardrails will decommission or restrict them due to quality or security incidents. Verification, the checklist says, should be proportional to the change's blast radius. m:…0730

Clay, a company that builds with agents daily, describes what that shift looks like inside a review process. It has moved the primary goal of code review from bug-finding to high-level architectural scrutiny — data schemas, service boundaries, system organisation — because, it says, the AI models it uses now rarely write bugs, and those that are written are typically caught by review agents. Complex architectural discussions now have to happen before a pull request is opened, and the developer experience team is building a PR bot to score the risk level of pull requests, with auto-approval for low-risk changes. The review meeting is no longer where bugs are found. It is where the architecture is argued about. m:…9465

The verification business

The market is paying for the gate rather than the generator. CodeRabbit says it has grown from 10 million to 100 million dollars in annual recurring revenue in just over a year, serving more than 20,000 customers and 400,000 monthly active users, by selling an AI-driven validation layer for software development — independent code review, security and triage. Its Deep Scans program, funded with 1 million dollars of free scans, surfaced more than 43,000 critical or high-severity findings in its first ten days, and the company joined the OpenAI Daybreak Defense Network in October 2026. The week's story, taken together, is not that agents can write code. It is that the industry's scarcest resource is now the check on the code, and the companies that sell the check are the ones compounding. m:…3584