Watch type: ignore in agent PRs, baseline your own DORA metrics, and review which internal services need a public IP
Two AI picks: Meta's Python typing survey on agents running type checkers and silencing them, and LeadDev on DORA pausing its annual survey. Two engineering picks: Cloudflare's GA of private origins behind its WAF, and Kubernetes AI Conformance becoming a SIG Architecture subproject.
AI & Agent Engineering
AI-assisted codingverified
Meta: in a 2026 Python typing survey, 24.4% had an AI coding tool run their type checker
Meta's 2026 Python Typing Survey had 1,794 responses. AI tools wrote type hints for 42.5% of 1,517 respondents, and 24.4% said an AI coding tool invoked their checker. Respondents also complained that agents spam type: ignore comments and use object for everything. Missing library type information was the top difficulty at 65.1%.
Why it matters
A type checker gives an agent a fast, objective signal, but only if the agent cannot silence it. This is a self-selected developer survey, and 29.8% of respondents were unsure types help AI-assisted coding.
Do this week
Count type: ignore additions in agent-authored pull requests over the last month. If they are rising, add a CI rule that fails new suppressions without a linked reason.
LeadDev: DORA will not run its annual survey this year, and AI productivity is still hard to tie to returns
LeadDev reports that DORA will not run its annual survey this year and is exploring more qualitative and observational research instead. It quotes Atlassian's State of Teams 2026 report: 89% of executives believe AI has increased the speed of work, but only 6% can connect that speed to organization-wide return on investment. Google has not explained the pause.
Why it matters
Many teams benchmark their delivery against the yearly DORA numbers. Without a new edition you must rely on your own data, and a vendor survey figure like the 89% and 6% split is not evidence about your team.
Do this week
Record your own four DORA metrics now as a baseline, and add one measure of rework or red builds so you can see whether agent-assisted work is raising speed or only churn.
Cloudflare: proxied DNS records can now point at private IPs, with WAF and bot management in front
Cloudflare made services for private origins generally available. A proxied A or AAAA record pointing at a private IP is reached over an existing Cloudflare One connection such as IPsec, GRE, Tunnel or Mesh, with WAF, bot management, rate limiting, caching and Workers applied. Argo Smart Routing and Dedicated CDN Egress IPs are not supported, and access needs the Private Origins entitlement.
Why it matters
Internal APIs, agent backends and MCP servers can take partner traffic without a public IP or an open firewall port. Cloudflare says a missing return route for its source ranges was the most common setup failure in the beta.
Do this week
List internal services that are exposed only to give partners or agents access. For one, check whether it could sit behind a private origin, and confirm the return route for Cloudflare's source ranges first.
Kubernetes: AI Conformance moves from a working group to a SIG Architecture subproject, with 31 certified platforms
The Kubernetes blog interviews Janet Kuo and Yuan Tang on AI Conformance moving from a working group to a SIG Architecture subproject. They say the program defined Kubernetes AI Requirements and grew certified platforms from 18 to 31 within a few months, with a roadmap toward automated conformance verification.
Why it matters
If you buy or run Kubernetes for GPU and inference workloads, a shared baseline reduces lock-in risk. The figures come from the program leads, and the certified list is the thing to check.
Do this week
When you next evaluate a managed Kubernetes platform for AI workloads, ask whether it is on the Certified Kubernetes AI Conformance list and which requirements it covers.
Count type ignore additions in agent-authored pull requests
Pick the last 30 days of pull requests with agent involvement. Count added lines containing type: ignore, noqa or eslint-disable, and compare with human-only pull requests of similar size.
Why. Agents can make a checker pass by silencing it, which hides the signal the checker is meant to give.
Measure. Suppressions added per 1,000 changed lines for agent versus human pull requests, and how many have a stated reason.
Classic / essential read
Sacrificial Architecture
Martin Fowler · 6 min
Fowler argues that code you will discard in a couple of years is not a failure, and uses eBay's rewrites and the rule of designing for 10x growth to make the case.
One lesson. Design for roughly 10x your current load and plan to replace the system before 100x, and say so openly so a rewrite is a planned step.