The one where we vibe code holiday cards | Season 5 Finale
MCP's "comfort zone," killing engineering dogmas, and the end of RTO mandates.
As the year draws to a close, the Dev Interrupted team reflects on a transformative year in engineering, spanning the rise of RAG to the emergence of agentic workflows. Then, we’re taking the conversation out of the booth and into the IDE.
Listen to Part 1 of our year-end wrap-up below. Once you’re done, head over to the Dev Interrupted YouTube channel for Part 2, where you can watch the team vibe code custom holiday cards and close out the year with some chaotic creativity.
1. MCP agents need real safeguards
Letting LLMs make tool calls is amazing, but without a decision-management layer, it is like hooking a nuclear reactor to your software. A new article from Adam Conway argues that the Model Context Protocol (MCP) lacks native support for transactional workflows, making it dangerous for irreversible actions like wiring money. MCP gives agents capability but not safety. The solution is to frame decisions on a spectrum of reversibility and consequence. If an action is low-consequence and highly reversible, it is in the “comfort zone.” If not, you need a human in the loop.
Read: It’s Not Too Late to Roll Back MCP
2. Your job is to deliver code, not just generate it
AI changes how we work, not who is accountable. A new piece from Simon Willison reinforces that software engineers cannot use AI generation as an excuse for buggy implementations. Just as a pilot is responsible for the plane even on autopilot, developers must manually and automatically verify every line of generated code. Velocity without proof simply shifts the burden onto reviewers and breaks trust. The bottom line is simple: regardless of who—or what—wrote the syntax, if you can’t prove it works, you didn’t ship it.
Read: Your job is to deliver code you have proven to work
3. Stop treating every PR like a moon launch
Rigid engineering dogmas are silently killing velocity. Anton Zaides challenges five “best practices” that have become bureaucratic bottlenecks, including the idea that every single code change requires a mandatory human review and the slavish adherence to 2-week sprints. He argues that modern teams should optimize for autonomy, allowing senior engineers to self-merge low-risk changes and moving away from processes that prioritize ritual over shipping. It is a call to audit your workflow: are your safeguards actually preventing bugs, or just preventing work?
Read: 5 engineering dogmas it’s time to retire
4. RTO mandates are a sunk-cost fallacy
Big Tech is souring on remote work, with Microsoft and others mandating three days in the office. However, this shift often looks like a sunk-cost problem rather than a productivity solution. While executives claim in-person work drives spontaneous collaboration, for many high-output individual contributors, that collaboration is often just an interruption. Mandates without incentives—like childcare or travel stipends—can feel less like culture building and more like a stealth layoff strategy.
Read: How big tech fell out of love with remote work
5. The internet grew 19% in 2025
Cloudflare’s Year in Review reveals that global internet traffic surged by nearly 20% in 2025, driven largely by the rise of AI agents and automated crawlers. Interestingly, 6.2% of that traffic was mitigated as malicious, highlighting the growing security cost of an automated web. But the most significant shift might be in defense. The share of traffic protected by post-quantum encryption nearly doubled to 52%, proving that the industry is aggressively fortifying itself against “harvest now, decrypt later” attacks before quantum computing becomes a widespread reality.
Read: 2025 Year in Review
6. The Dr. Seuss guide to agentic tools
For a recent internal talk, I leaned into an old teaching tactic: make it fun to make it stick. By pairing Claude with Nano Banana for visuals, I created a Dr. Seuss-style picture book called 'Oh, the Tools You’ll Connect!' It turns out the best way to explain agentic tool-use isn’t a white paper, but a story about Claude living in a box. It’s a lesson in AI enablement, proving that high-friction concepts stick better when you lower the stakes.
Read: Oh, the Tools You’ll Connect!
