Post

Module 8 — Bedrock resources in Terraform

Posted
By
2 min read

Prompt & Tool Design for .NET Teams · Part 3 — Infrastructure as Code · Module 8 of 12

A Bedrock agent you created by clicking around the console is an agent nobody can reproduce and everybody is slightly afraid of. This module declares the same thing in Terraform, where it can be reviewed, diffed, and destroyed on purpose. If you’ve written Terraform for anything else, none of this is new. It’s just new resource names.

Objective

Define a Bedrock managed agent, an action group (the set of tools the agent is allowed to call), and a guardrail (a safety filter on what goes in and what comes out) — all declaratively.

Read (~10 min)

Lab (~20 min)

Start with the agent and its execution role. This is the verified shape from the current provider — note foundation_model reads from a variable, which is the whole point of the closing exercise:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

variable "model_id" {
  type    = string
  default = "anthropic.claude-3-5-haiku-20241022-v1:0"
}

data "aws_iam_policy_document" "agent_trust" {
  statement {
    actions = ["sts:AssumeRole"]
    principals {
      type        = "Service"
      identifiers = ["bedrock.amazonaws.com"]
    }
  }
}

resource "aws_iam_role" "agent" {
  name_prefix        = "AmazonBedrockExecutionRoleForAgents_"
  assume_role_policy = data.aws_iam_policy_document.agent_trust.json
}

resource "aws_bedrockagent_agent" "comic_assistant" {
  agent_name                  = "comic-assistant"
  agent_resource_role_arn     = aws_iam_role.agent.arn
  foundation_model            = var.model_id
  idle_session_ttl_in_seconds = 500
  # Your system prompt — keep it short and specific.
  instruction                 = "You are a witty assistant that recommends programming comics."
}

Then add, following the walkthrough above:

  • an action group pointing at a stub Lambda — a C# Lambda that reuses your ComicPlugin logic;
  • an aws_bedrock_guardrail with a basic personally identifiable information (PII) filter;
  • the IAM permissions the agent needs to invoke the model.

Test it in the console’s agent test pane, then terraform destroy when you’re done. (Do the destroy. An idle agent costs nothing, but tidy is a habit.)

Done when

terraform apply gives you a working agent, and the foundation model is a var.model_id. So a model upgrade is a one-line change to your tfvars, not a hunt through the codebase.

Series navigation

Tutorial, AI
artificial-intelligence infrastructure-as-code
This post is licensed under CC BY 4.0 by the author.