The middle should be boring
A new pre-alpha project called Neckbeard has the right instinct for agent-built infrastructure. It asks the agent to inspect an application and explain choices. It moves planning and rendering into deterministic code backed by a versioned catalog. Human review stays in the route.
The useful idea is not that an agent can write HCL. It is that the agent should have fewer chances to improvise it.
The repository labels its current boundary. Static validation, authenticated planning, and deployment checks have separate verdicts. Its README says the deployment lifecycle is experimental. That is better than a green terminal block that quietly promotes a syntax check into a production claim.
Validation has a hard edge
OpenTofu documents that tofu validate checks syntax and internal consistency without remote state or provider APIs. Checks that need state are skipped. Use it as one gate, then stop the claim there.
Cost output needs the same restraint. Infracost separates baseline costs from usage costs. Usage estimates depend on monthly values supplied by the operator or a central policy. An attractive monthly total without those assumptions is a decorated guess.
Security is machinery
A current CSO report quotes Gary McGraw, Rob T. Lee, and Dan Wallach arguing that recent agent incidents point to weak containment and monitoring. Their advice is plain: isolate the process, control egress, watch behavior, and keep normal security ownership in the room.
The MCP security guide reaches the same shop floor. It calls for exact command display before local server execution, explicit consent, restricted file and network access, and minimal scopes. The model can propose work. The host still owns authority.