AI
How IaC Became the Default
Brendan Burns gave a keynote in 2012 arguing that infrastructure as code was worth the effort. By 2025, nobody needs that argument. Cloud turned infrastructure into an API. Rising reliability expectations made every internet-facing service operate like a bank. Infrastructure as code became as obvious as unit testing.
Adobe’s story was messier. Tyler Jacobsen inherited a company built from 28 acquisitions, each with its own stack, producing what he called “relative chaos” and click ops everywhere. Standardizing on reusable Terraform modules got 900 teams onboarded and cut outages from manual provisioning errors.
The Context Gap in AI-Driven Infrastructure
AI applied to infrastructure today is mostly context-free. Burns put it plainly:
“it’s kind of like hiring the world’s greatest s who knows precisely nothing about how your company operates” (Brendan Burns)
Microsoft’s Azure resource graph encodes those relationships (DNS to IP, server to database, VM to on-call rotation) so automation contacts the right team when a CVE appears, not just logs it.
Adobe’s Terraform MCP server lets engineers state intent in natural language and reach 90-95% of the output. But Jacobsen flagged the tradeoff: less-experienced engineers writing IaC are exactly when policy-as-code matters most. Armon estimated 30-40% of Terraform is already AI-generated.
Three Layers of an Agentic Operating Model
Agentic infrastructure breaks into three unsolved layers. Runtime: AI workloads are already striping across clouds or moving to private GPU farms because capacity and data-sensitivity constraints rule out single-provider deployments. Gateway: agents need the same thing microservices needed: a registry, routing, and policy enforcement for agent-to-agent calls, whether the protocol is MCP or A2A. Data: most large enterprises run hundreds of systems of record, each with its own access controls and query interface. Until those are accessible to agents in a consistent way, context stays low.
Project Infograph: A Graph-Based Infrastructure View
Project Infograph is HashiCorp’s answer to the data fragmentation problem: a graph database built for multicloud estates. It connects Terraform, HCP Packer, and AWS today, with Azure, Google Cloud, and Wiz planned. The graph stores relationships, not just inventory. It knows which workspace manages which EC2 instance and which Packer artifact produced which AMI.
Jackie walked through the Project Infograph live demo (32:37) on stage. Pre-built queries surface EC2 instances older than 30 days, infrastructure on outdated Terraform versions, and images not managed by Packer. The graph view shows the full blast radius of a version update before anything is changed.
Identity for Agents, Not Just Humans
“joke that the S in AI is for security.” (Armon Dadgar)
Identity was designed for humans. Agents break every assumption: a credential checked out by a person becomes a query run autonomously by an agent, and the ticket-driven onboarding process that works at human scale collapses under thousands of service principals. Dadgar mapped the spectrum from workforce SSO through privileged access management, secrets, certificate lifecycle, and non-human identity. HashiCorp covers the middle with Vault and Boundary. IBM Verify covers the human end. MCP interfaces for Vault, Radar, and Consul are already shipping to make those products agent-ready.
Notable Quotes
it’s kind of like hiring the world’s greatest s who knows precisely nothing about how your company operates Brendan Burns · ▶ 8:41
creativity and consistency rarely converge Tyler Jacobsen · ▶ 12:38
it was relative chaos. Tyler Jacobsen · ▶ 13:20
joke that the S in AI is for security. Armon Dadgar · ▶ 37:37
Key Takeaways
- AI agents are context-free today. They need relationship graphs, not just inventory, to reason about infrastructure.
- Adobe onboarded 900 teams by abstracting Terraform and Vault behind reusable modules and secure-by-default patterns.
- Project Infograph gives a single resource graph across AWS, Terraform, and Packer, cutting multi-day audits to minutes.