AI agents

AI agents and infrastructure as code fit together unusually well. An agent working on a Terraform codebase produces a plan before anything changes, and that plan is a review gate: you read exactly what will happen before it happens. Pull requests add a second gate on the code itself. Between a probabilistic model and a live AWS account sit two deterministic checkpoints, a diff and an approval. That determinism is the bridge. A model can be confidently wrong, and the plan makes the wrongness visible while it is still cheap to fix. Most software work gives agents no comparable safety net. Infrastructure work had one built in years before agents arrived.

>>> Before deploying agents that touch production, make sure you start with human in the loop approval flows and promote autonomy as the agent earns your trust. Track every agent decision, prompt trace, and proposed action in a read-only ledger. Restrict permissions to isolated sandboxes or specific resource tags until confidence is proven. Implement an immediate global override to revoke agent credentials and halt execution instantly.

That intersection is the thesis of this whole site, which makes this page its flagship hub. I work on it daily as a cloud infrastructure architect and AWS re:Invent speaker, and most of my recent writing circles back to it. The collection below mixes two formats. Original posts live here in full. Link posts carry short commentary on articles published with partners such as Spacelift, Gruntwork, Qodo, StackGuardian, and AWS. Each commentary explains why the piece matters and who should read it, so you can decide in half a minute whether the full article deserves your time.

The sections are ordered the way I would read them. Agents meet infrastructure as code is the core: Claude Code against Terraform, layered guardrails for agents that touch infrastructure, MCP as the protocol that gives them tools, and an agent-driven migration on AWS. Assistants in the SDLC steps back to the wider development loop, where test-driven development and coverage analysis give generated code the same checkable targets that plans give infrastructure. The platform angle covers what all of this does to platform engineering, automation maturity, and team skills. If you only have time for one post, start with the Claude Code guide. It is the closest thing to a summary of everything else on this page.

A note on shelf life. The tooling in these posts moves fast, and specific product names will age. The underlying pattern ages slower: give the agent a constrained workspace, force every change through plan and review, and keep credentials out of its reach. That pattern holds whether the assistant is Claude Code, Kiro CLI, or whatever ships next quarter. Everything here is tagged, so new posts land on the ai-agents tag page as soon as they publish, the full feed mixes in the rest of my writing, and the newsletter covers the same ground every two weeks.

>>> I expect agents in the near future to fully own routine, deterministic infra tasks such as automated deployments, baseline anomaly triage, standard incident remediation, and capacity scaling. Humans will retain ultimate ownership of high-risk architectural strategy, complex cross-system debugging, policy governance, and ethical or strategic budget decisions requiring nuanced business context.