The community had a lot to say this week, and most of it was pointing at the same problem from different angles. AI review loops that never terminate. Datadog bills that keep climbing. Jenkins still running in shops that swear they've moved on. Entry-level candidates hitting walls. What ties all of this together is a discipline gap: the tools are more powerful than the processes surrounding them, and most teams haven't caught up yet. That's what this issue is about.

One hiring manager posted that 5,000 people applied to his last opening. Five thousand. That number alone explains why entry-level DevOps candidates are getting zero callbacks after months of applying. The market isn't broken for everyone, but it is brutally competitive for anyone without proven production experience.

The roles with traction right now are SRE, Platform Engineer, and DevSecOps. AWS cloud depth and multi-cloud IaC skills are still pulling weight. AI Ops is showing up in job descriptions with more regularity, though what that actually means varies wildly from employer to employer.

The advice that keeps surfacing from practitioners: shadow seniors, show up with something you've already tried, and stop waiting to be handed a curriculum. Referrals still move the needle more than any cert. If you're early in your career and you have a manager who will let you sit in on incident reviews, take every single one.

The AI review loop problem got more airtime this week than any other AI topic, and it's worth understanding what's actually happening. You set up a feedback loop: Claude reviews your code, Codex suggests changes, Claude reviews again. The loop never terminates because the acceptance criteria aren't machine-gradable. Opinions don't run out. Style preferences compound. You end up with fifty iterations of nitpicking and nothing shipped.

The fix isn't technical. It's process. Define a hard checklist upfront: security scan passes, test coverage threshold met, no new linting errors. Once the code clears the checklist, stop asking the model for more feedback. The model will always find something. That is its job. Your job is to have a definition of done that doesn't bend.

On the tooling front, Cursor's Composer 2.5 is drawing strong reactions for day-to-day codegen. Claude Code continues to lead on planning and large-scope work. The practitioners getting the most out of these tools are pairing them with local markdown wikis, essentially giving the model persistent context about the codebase, the conventions, and the decisions already made. One benchmark put a well-structured markdown wiki above every other memory approach tested.

A few projects worth bookmarking this week. Hull is positioning itself as a Helm alternative with a "YAML-by-construction" approach: multi-environment support, signed charts, RBAC, and delta sync baked in. If you've been fighting Helm's templating ceiling, it's worth a look.

InfraLens uses eBPF for live service-mesh topology visualization with no sidecars required. That's the part that matters: no sidecars means no performance overhead and no configuration gymnastics to get it running.

Whatbroke snapshots dependencies, config, environment variables, and Dockerfiles so you can diff against a known-good state when something breaks. Simple concept, genuinely useful when you're trying to figure out what changed between the last green build and now. There's also Understudy, which targets high availability for spot-instance single-replica services on Kubernetes, giving near-zero downtime without requiring you to run multiple replicas.

The strongest practical takeaway from CI/CD discussion this week: move logic out of YAML and into tested shell scripts or Makefiles. Call a make target from your pipeline. Pipeline YAML handles the trigger and the environment. The logic lives somewhere you can actually test it.

IAM is the leading cause of gray hair this week. Least-privilege policies on AWS are theoretically straightforward and practically exhausting. The pattern that's getting traction: let developers run wide in a dev account, then point IAM Access Analyzer at CloudTrail and generate the scoped policy from what the role actually called. You get a policy grounded in real usage instead of one you assembled from documentation and guesswork.

Observability costs continue to be a pressure point. One team at 300GB/day of logs on Datadog described the bill curve as going vertical. The migration math to Loki and Mimir looks compelling until you account for the engineer time required to run it well. The teams who've actually moved their bill down say the first win came from rationalization at ingestion, not from swapping backends. Cutting log noise before it hits the pipeline saves more than any platform migration.

Blue/green deployments remain a recurring source of confusion. They work cleanly for stateless services. They do not work cleanly for shared databases unless your migrations are backward compatible and your rollback plan accounts for that. If they're not, rolling updates are almost always the right call.

A 150-comment thread titled "Anyone still using Jenkins?" is a decent proxy for where the community actually stands on something. The answer: yes, many shops are still running it, and that's not going to change soon.

The criticism is familiar: plugin dependency fragility, fat pet controllers with 80 plugins, pipeline config living in the UI instead of in source control. The people defending it make a fair point: nothing else offers the same flexibility at zero licensing cost. For shops with hundreds of pipelines, the migration calculus usually doesn't work unless someone is in real pain or has an executive sponsor willing to own the timeline.

The practical advice for shops still running Jenkins: ephemeral agents, JCasC (Jenkins Configuration as Code), and a hard limit on plugin count. Those three changes buy more than a grand migration plan. Most of the Jenkins pain is self-inflicted by the way the tool is being run, not the tool itself.

Verdict: don't migrate unless you're genuinely in pain. Do fix your Jenkins setup. Those are different problems.

AWS added real-time vector search to DynamoDB. The technical notes are interesting: the index mechanism is getting scrutiny on cost and query performance at scale. Separately, Bedrock AgentCore and the Strands framework are drawing practitioners who want to understand what's actually underneath the abstracted surface before committing to it. Early reports suggest the architecture is shifting fast enough that getting hands-on now is worth the time.

Compartment did a full architecture rewrite of their open-source internal app platform. Worth reading for the design decisions, not just the announcement. Hull (Helm alternative), InfraLens (eBPF topology), and similar FOSS tools are getting community energy as teams look to cut managed platform costs. The Cursor team plan pricing discussion also generated significant heat this week: people are doing the math on the token markup and moving plans.

Open-weight LLMs (DeepSeek V4 Flash, Kimi K3, Qwen3.5) are showing up in agentic workflow benchmarks with increasing frequency. The gap between frontier closed models and the best open-weight options continues to close on specific task categories.

The Weekly Draft

Thanks for reading. If this landed in your inbox, forward it to someone who'd find it useful. If you're reading it in the archive, the subscribe link is below.

You're receiving this because you signed up for The Draft. Unsubscribe anytime.