AI
Data Domain Sprawl and Organizational Structure for AI
Asking an AI which apps are impacted by Log4j requires data from GitHub, Artifactory, Snyk, Terraform, Ansible, Wiz, and mainframe. That is 55 systems for one question. Infrastructure as code already forced enterprises through this shift once, from vertical technology silos (a VMware team, an F5 team, a Palo team) to horizontal workflow ownership. AI forces the same shift again, but for data. The answer is data domains: InfraGraph consolidates all infrastructure data into one queryable domain, the way a platform team owns all infrastructure provisioning.
Brownfield Enterprises: Federation Over Consolidation
Existing enterprises will not consolidate 100 systems down to 10. Dadgar’s read: standardize on protocols, then federate. A2A and MCP are already pointing toward that model. North-south and east-west gateways connect the agents representing those 100 systems. A service registry ties them together. Consul already handles this for some customers. New companies can start with fewer, more automatable platforms. Legacy enterprises cannot throw away decades of infrastructure, so the protocol layer becomes the integration surface rather than a full replacement.
Agents as Microservices: Applying Existing Infrastructure Patterns
Agents running as root in production with no monitoring are not a novel problem. Dadgar called it Groundhog’s Day: the industry keeps asking how to track which agents run in an environment, and the answer is the same one microservices required. Register them in Consul. They speak HTTP. They need identity, governance, observability, and a service mesh. The specific nuances differ, but the infrastructure checklist maps almost exactly to what distributed services already demand.
LLM Training Data Advantages Reinforce Incumbent Tools
LLM training data already picks technology winners. Python is displacing R, Stata, and MATLAB in data science because models trained on more Python code simply perform better on it. HCL generates better Terraform than CDK Python for the same reason: more training examples exist. Dadgar compared it to Google’s monthly index updates. In the early days you could tell when the new index rolled out. Eventually Google moved to continuous indexing. Foundation models will reach that point too, and the new-framework penalty will shrink. For now, incumbents with large training corpora have a measurable edge.
Three Unsolved Layers for Safe Agentic AI
Three layers need to settle before agentic AI runs safely at scale. Runtime: containers and Kubernetes may win again, but serverless and WebAssembly are live contenders. Red Hat is building vLLM; no clear winner yet. Gateway: A2A and MCP are frontrunners, but both lack native request tracing and user identity propagation down the call chain. Extensions are already planned for MCP. Data: Dadgar called this the hardest layer. Connecting Salesforce to SAP requires those vendors to open their platforms. Neither has. Until standard query interfaces emerge, the tooling cannot be built.
Q&A
How can InfraGraph help enterprises understand common Terraform patterns across 200 app teams? InfraGraph pulls all 200 Terraform workspaces into a graph database where you can query for common patterns and ask an agent to identify the top 50 modules your estate actually needs. ▶ 9:15
Does HashiCorp plan to expand beyond infrastructure into application pipelines with the new Actions feature? HashiCorp stays focused on infrastructure and security; CI/CD lives in GitHub and GitLab, but the CD boundary blurs where Terraform actions handle production-touching changes with full audit trails and policy as code. ▶ 12:18
How should a consultant convince a large healthcare client to use Terraform over OpenTofu when one developer has decided otherwise? Build a coalition above the individual: reach leadership who weigh talent pipeline, bus factor, vendor support, and enterprise scalability rather than personal tool preference. ▶ 22:10
What is the one lasting piece of advice for HashiConf 2025 attendees? Optimize for workflow problems, not technology choices: teams that chased vector databases found them irrelevant within one product cycle when everything moved to agents and PGVector. ▶ 27:37
Notable Quotes
I’m really using AI as the glue layer Armon Dadgar · ▶ 5:13
well, it deleted your database. That Armon Dadgar · ▶ 7:40
it’s like the LMS takes so long to train Armon Dadgar · ▶ 17:33
Key Takeaways
- AI agents face the same 55-system data sprawl problem that IaC solved for workflows.
- Brownfield enterprises should federate 100 systems via A2A and MCP, not consolidate them.
- Workflows and fundamentals outlast technology choices: vector databases became irrelevant inside one product cycle.