*Green is Swamp. Purple is Stategraph. Orange is the expected position in 2 to 3 years. A white bar is inertia. All positions are AI judgement, i.e. the industry median.*

## The read

Nobody switches IaC off. Its layers lose value one at a time, and the installed base keeps it running for years. [Adam Jacob says he is never using traditional IaC again, and Darren Shepherd says agents will maintain Terraform and CDK just fine](https://teapawt.org/adamhjk). They argue about different layers of this map.

[Sean goes further](https://teapawt.org/webframp/status/2117090770415169859#m):

> IaC isn’t dying it’s already dead and was always a dead end we just couldn’t see it before. And now attachment to it’s comfortable lie of predictability has people stuck using poor tools to solve modern problems

1. **Authoring gets cheap before it gets reliable.** Agents write HCL at low cost, and both Adam and Darren accept this. Agents wrote about 23% of public IaC pull requests in May 2026 ([Barrak et al.](https://arxiv.org/html/2610.34613v1)). Single-attempt pass rates on one Terraform benchmark range from 14.0% to 62.4% by model ([arXiv 2607.20478](https://arxiv.org/html/2607.20478)). It still removes the main reason to design a tool around a language that humans write.
2. **The engine is becoming interchangeable.** Terraform and OpenTofu share a language and providers. [Pulumi now runs HCL](https://www.pulumi.com/blog/all-iac-including-terraform-and-hcl/) and has a Terraform state backend. [Stategraph](https://github.com/stategraph/stategraph) runs both binaries. When vendors offer the rival interface, the interface is the standard and the engine stops being the differentiator.
3. **Day-2 and app + infra work is still custom.** This is the gap Adam names. In [Firefly's vendor survey](https://www.firefly.ai/state-of-iac-2026), 90% say their IaC orchestration falls short, and about one third tie drift to a costly production incident.
4. **Agent autonomy is limited by guardrails.** In the same survey, 34% would trust agents with autonomous production changes, and 42% name missing guardrails as the top blocker. Swamp's fixed behaviors and vaults answer this blocker. So do the [Terraform MCP server](https://www.hashicorp.com/en/blog/terraform-mcp-server-is-now-generally-available) and Spacelift's governed access.
5. **Two parts of the state layer are still custom.** Run history with typed data records what was done. A resource graph records what exists and what depends on what. The first is in Swamp, the second is in Stategraph. They are complements.
6. **Incumbents are also moving up.** Pulumi Neo, Terraform Actions, [Infragraph and the Terraform Developer Portal](https://www.ibm.com/new/announcements/terraform-advances-infrastructure-lifecycle-management-for-the-agentic-era) all target this space. Swamp's window is real and contested.