I’ll read the article, identify its gaps, and check current evidence for how much progress has been made. I’m using the agent-reach skill for the research.

The article mixes gaps we can address with engineering today—environments, checks, and failure reporting—with harder questions about whether those checks capture everything that matters. I’m checking that distinction against production reports. Agent-reach’s search backend isn’t configured here, so I’m using web search.

We can address most of the operational gaps today. Reliably closing all of them—enough to leave arbitrary production software entirely unattended—is still an open problem with no defensible completion date.

Assuming “we” means the industry, that’s my assessment as of September 12, 2026. The article was updated September 5, so its concerns are current.

The important distinction is between having a mitigation and having enough evidence to remove human oversight.

Three pieces of evidence explain why I draw that line:

• Autonomous implementation already exists in bounded settings. StrongDM reports generating and validating software without human code review, but humans still define intent, scenarios, and constraints. That demonstrates a particular operating model, not universal autonomy. StrongDM’s account
• Independent verification is improving, but remains fallible. Anthropic describes combining executable checks, model graders, and human evaluation. The key is evidence about the outcome—not an agent’s declaration that it finished. Evaluation guidance
• Difficult maintenance still exposes major gaps. An August research preprint evaluating whole-repository migrations reports that only 28 of 520 runs passed all three evaluation stages. This is a demanding, narrow benchmark, not a general coding success rate, but it directly challenges the idea that reliable long-term maintenance is solved. SWE Refactor Bench

I’d also qualify two claims in the article. Using a different reviewer model does not guarantee independent judgment. And hiding secrets does not automatically contain every attack. Anthropic’s September incident assessment explicitly says evaluations that reliably expose every deployment failure remain unsolved. Those incidents occurred in cybersecurity evaluations, so they shouldn’t be treated as ordinary coding failure rates. September assessment

For planning, my estimates—not measured industry timelines—would be:

• Weeks to a few months: a useful supervised factory for a well-tested repository.
• Several months or longer: selective unattended shipping, after accumulating evidence for specific low-risk changes.
• No credible date: general unattended development and maintenance across arbitrary production systems.

The milestone I’d use is “this class of changes ships with acceptable defects, rework, and cost over repeated releases.” That’s a stronger basis for increasing autonomy than either model benchmark scores or the number of generated PRs.