Pulumi Redesigns Agent Architecture to Tackle Workspace Continuity Over Git State

Pulumi published an architectural deep dive detailing how it solved the workspace recovery bottleneck in Pulumi Neo, its AI infrastructure agent. Initially, hosted agent executions relied on Git primitives—persisting commit hashes, branch pointers, and local diffs—to reconstruct the environment after runtime interruptions. However, this approach failed across multi-step execution graphs: large generated SDKs exceeded payload limits, intermediate build artifacts were lost, and tasks failed if remote repositories were altered or deleted mid-run. To fix this, Pulumi decoupled conversation replay from filesystem persistence, adopting Kopia-backed incremental, encrypted snapshots stored directly in task-scoped object storage. This engineering redesign highlights a pivotal distinction in autonomous DevOps tooling: conversational context is not equal to operational continuity. An infrastructure agent does not simply generate declarative manifests; it runs dependency installations, evaluates resource previews, generates SDK stubs, and troubleshoots runtime validation failures across multiple execution steps. When an execution runtime restarts due to cold starts or task migration across managed worker processes, restoring only the prompt history leaves the agent without its actual workspace state, causing downstream failures. This move fits into the broader evolution of AI-driven platform engineering and infrastructure management. As cloud teams transition from reactive code completion (such as GitHub Copilot) to long-running autonomous workflows that span discovery, provisioning, and policy remediation, agent environments must be treated as stateful, ephemeral compute workloads. Isolating the control plane (which issues scoped, temporary credentials) from the data plane (which streams encrypted filesystem chunks) allows platforms to provide robust crash recovery while maintaining strict multi-tenant boundary isolation. For platform and DevOps engineers integrating AI agents into continuous delivery pipelines, this architecture offers key lessons. Treating version control systems like Git as state caches introduces severe fragility when handling binary artifacts, intermediate dependencies, and dynamic caches. Teams building custom agentic workflows should separate conversational memory from filesystem state, implement lease-based worker scheduling to prevent split-brain updates, and ensure snapshotting processes operate asynchronously to avoid penalizing agent execution latency.