Blog: Your Coding Agents Can Already Build Infrastructure. That's the Problem.

[dirien]

What this is

A perspectives post by @dirien that answers the question sales keeps hearing from VPs of engineering at AI startups: "Why can't Cursor/Claude and my engineers just do this?"

The answer the post makes: they can, fully conceded up front. What they can't be is their own control plane. The gap between AI adoption (88% of platform engineers daily) and AI returns (5.5% of companies) is a control problem, not a capability problem, and the platform underneath the agents is the durable decision.

Structure

1. Full concession (coding agents do real IaC work; teams already use them)
2. The adoption/returns gap, with a different-surveys honesty hedge (McKinsey, MIT, Gartner)
3. What breaks at agent speed: review, credentials, policy, audit, anchored by the July 2025 Replit and Amazon Q incidents and Snyk's ToxicSkills audit
4. The control plane (preview, policy as code, task-scoped credentials, named approvals, audit), including the build-vs-buy follow-up (GitHub/OPA/Vault/CloudTrail)
5. Separation-of-duties answer to the literal question, with an arbitration example
6. Pulumi mapped in once: Agent Skills + MCP server for BYO agents, Neo as one governed agent, IDP as the paved road
7. A five-question vendor-neutral scorecard for the buying conversation

Review status

• Every statistic and incident verified against primary or near-primary sources; all external links return 200
• make lint passes; docs-review and humanizer passes applied (zero em dashes, no AI-pattern vocabulary)
• Feature image rendered from the shield template via /blog-feature-image; social copy for X/LinkedIn/Bluesky in frontmatter
• Positioning follows the AI/agent spectrum rules: Claude Code, Cursor, Codex, and Neo all ride the same governed path

🤖 Generated with Claude Code

[github-actions]

Social Media Review

content/blog/coding-agents-can-build-infrastructure-thats-the-problem/index.md

X — FAIL

• Over character limit: 318 chars vs. 255 limit
• First-person voice: "asked me" / "we were talking" has no named subject (corporate account can't speak as "me") — should attribute to Engin Diri like the LinkedIn copy does

LinkedIn — PASS

Bluesky — PASS

---

Suggested copy

X (245/255 chars) — minimum-change repair: attributed the quote to Engin Diri (fixes voice) and trimmed the second paragraph's tail to fit the limit:

"Why can't Claude and my engineers just do this?" A VP of engineering asked Engin Diri that during a conversation about coding agents building infrastructure.

The honest answer: they can. If that's where your thinking stops, that's the problem.

---

Suggestions (advisory)

LinkedIn

• Solid as-is — named attribution, concrete stats, and a clear pointer to what Engin covers in the post.

Bluesky

• Solid as-is — withholds the "control plane" answer and ends on the curiosity gap.

---

To apply these suggestions, comment: @claude please update the social posts in the frontmatter with the suggested copy from the social media review above

To re-run the social media review after updates, comment: /social-review

Updated for commit 00f6c1ba89946174b1c7fb864f4b05090906e6f0 (short: 00f6c1b) at 2026-07-23 15:15 UTC.

[pulumi-bot]

Your site preview for commit 33b5aec is ready! 🎉

http://www-testing-pulumi-docs-origin-pr-20471-33b5aec0.s3-website.us-west-2.amazonaws.com

Changed pages:

• /blog/coding-agents-can-build-infrastructure-thats-the-problem/