Terraform changes with the decision attached

DevOps, SRE and securityPlatform team changing infrastructure with Claude Code

We moved our database off Aurora Serverless to a fixed r6g.large because scaling pauses broke checkout at peak. Three months later an agent, asked to 'cut database cost', proposed going back to serverless in a tidy Terraform plan. It had no idea why we left.

How do I stop an AI agent undoing an infrastructure decision we made for a good reason?

Save the choice in ArcRoom with arcroom decide, with the reason and the incident it came from. Recall then puts the decision in front of any agent working in that Terraform module, and a rule on the module can say what lives there, like 'fixed size on purpose'. ArcRoom records terraform plan and apply commands, never the plan output or state. ArcRoom can't block a terraform apply.

Also asked as: An agent proposed going back to a setup we left after an outage. How do I prevent that? · How do I give Claude Code the history behind our Terraform choices? · Can an AI agent see why our database is sized the way it is?

The work

The choice is saved with arcroom decide, with the reason and the incident it came from.

ArcRoom keeps
  • Decision
  • Recall
The next agent gets

Recall (budget 1.0k tokens):

Step by step

  1. The choice is saved with arcroom decide, with the reason and the incident it came from.
  2. ArcRoom records terraform plan and apply commands and the .tf files touched, never the plan output or state.
  3. Recall puts the decision in front of any agent working in the database module.
  4. A rule on the folder says what lives there ('modules/db: prod RDS, fixed size on purpose').

What the agent sees

Recall (budget 1.0k tokens):
• Decision (Jul 22): prod DB stays provisioned r6g.large, not serverless.
  Why: scaling pauses caused checkout timeouts at peak (incident #41)
• Code map: modules/db: prod RDS, sized by hand

Compared with today's tools

ADRs in the repo + CLAUDE.md

Write an architecture decision record per choice and point CLAUDE.md at the folder.

Setup
1 hour
Ongoing
write an ADR per decision, ~15 min; agents must be told to read them

ArcRoom

Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.

Setup
2 min (npx arcroom init)
Ongoing
None

ADRs are the better long-form record; ArcRoom brings the one relevant decision to the agent at the moment it edits that module.

Where it stops

ArcRoom tells the agent; it can't block a terraform apply. A policy layer agents can't bypass is on the roadmap.

Try it in your project

Works with the agents your team already uses, and any agent over MCP or the HTTP API.

More in devops, sre and security