May 2026 Summaries

Pulumi has introduced version 1.0 of the Pulumi Service Provider, marking a significant advancement in using Pulumi itself to manage Pulumi Cloud. This update leverages the Pulumi Cloud OpenAPI specification, allowing new features and resources to be available in the provider simultaneously with their release in Pulumi Cloud. Key enhancements include fine-grained role-based access control (RBAC) as code, Pulumi Identity Provider (IDP) as code, and audit log export as infrastructure as code (IaC). These improvements aim to streamline infrastructure management and automation by making new capabilities immediately accessible across all supported languages. The update also maintains backward compatibility with existing resources, and feedback during its preview phase is encouraged to refine the offering.

Anthropic's piece "How Claude Code Works in Large Codebases" emphasizes that the surrounding ecosystem, or "harness," is more crucial to the performance of coding agents like Claude Code than the model itself. The harness, comprised of seven components, provides context and tools that tailor the agent’s operation within a specific code repository, eliminating the need for a codebase index. This setup allows the agent to navigate using basic shell tools, though it initially lacks a comprehensive understanding without contextual guidance. Key elements include a foundational CLAUDE.md file, hooks for self-improvement, path-scoped skills, and symbol-level search capabilities via language and MCP servers, which enhance navigation and reduce brute-force searching. Additionally, subagents are employed to separate exploration from editing, preserving context and efficiency, while periodic reviews and updates to the harness ensure ongoing relevance and performance improvement. The success of implementing Claude Code at scale requires dedicated ownership and strategic integration to prevent fragmentation and ensure consistent use across teams.

AI infrastructure is bifurcating into two main domains: one focusing on the physical resources like GPUs, schedulers, and MLOps platforms for running AI workloads, and the other on AI systems that manage infrastructure by automating tasks such as generating, deploying, and governing cloud resources. As the demand for both types of AI infrastructure grows, organizations face the dual challenge of investing in compute resources and adopting AI-powered management tools. McKinsey research highlights a significant productivity boost from generative AI, driving platform teams to enhance infrastructure capabilities to support increased application development. The guide explores various tools across these domains, including specialized GPU clouds like CoreWeave and Lambda Labs, serverless GPU solutions such as Modal, and MLOps platforms like Weights & Biases and MLflow, emphasizing the scalability and efficiency they bring to AI operations. Additionally, it delves into AI-powered infrastructure management solutions, such as Pulumi Neo, Firefly AIaC, and Spacelift AI, which offer automated code generation and execution capabilities to streamline infrastructure governance and compliance. The strategic selection of tools depends largely on an organization's existing cloud strategy, team expertise, compliance requirements, and budget, with a trend towards integrating AI as a multiplier for routine tasks while retaining human oversight for complex decision-making.

Pulumi has introduced "pulumi do," a command designed for quick, direct interactions with cloud resources without the need for extensive infrastructure as code (IaC) setups. This tool simplifies operations like creating, reading, updating, or deleting cloud resources across various providers by allowing users to execute commands directly from the terminal. It leverages the Pulumi engine and resource model to streamline the process, reducing the need for multiple setup steps typically required in traditional IaC workflows. "Pulumi do" is particularly beneficial for ad-hoc tasks, enabling both humans and AI agents to manage infrastructure efficiently without in-depth coding knowledge. It supports a consistent command structure, predictable output, and works across multiple cloud providers, minimizing the need for provider-specific tooling. Future enhancements include unified credential management, cross-resource references, and a stateful mode that will allow for drift detection and lifecycle management, paving the way for a seamless transition to full IaC when needed. This new feature is available as a research preview in Pulumi CLI version 3.242.0 and later, and it invites user feedback to refine its capabilities.

Ewan Dawson, CTO of Compostable AI, discusses the transformation of software development into an AI-native factory model, emphasizing the importance of adopting industrial processes to maximize efficiency and value delivery. He argues that traditional software development, seen as a craft, must evolve into a more automated and scalable framework, leveraging AI to reframe and eliminate problems rather than simply solving them. Dawson shares insights from his experience building an AI-native engineering organization, highlighting the need for specialized agents and tools like Pulumi to manage infrastructure efficiently. He underscores the importance of designing for convergence rather than one-shot correctness, and the necessity of running these automated processes in a cloud environment to maintain security and facilitate continuous iteration. Dawson concludes that the future of software development lies in identifying and updating outdated practices to fully harness the benefits of AI-driven technology.

Pulumi Neo introduces a new feature allowing users to automate platform maintenance tasks by setting custom cadences for running tasks and automatically opening pull requests for each run, addressing the challenge of keeping provider versions up-to-date across stacks. This automation can include built-in templates like provider freshness checks, encryption audits, and activity digests, or users can create custom prompts with flexible scheduling options. Automations operate with default settings for safety, such as automatic approval and read-only permission modes, which can be adjusted as needed. These tasks run in the same context as interactive tasks, adhering to established naming conventions, tagging policies, and architecture rules, while using the RBAC permissions of the user who scheduled them. Pulumi encourages users to explore these features through the Automations tab in Pulumi Cloud and invites feedback to enhance the system further.

Pulumi Neo has expanded its functionality to integrate with GitHub and Slack, allowing users to involve the agent in discussions and tasks directly within these platforms. By mentioning @pulumi-neo in GitHub pull requests or issues, or @Neo in Slack channels, users can initiate tasks, such as reviewing changes or investigating issues, with Neo providing detailed responses within the same thread. This functionality enables seamless collaboration across platforms, ensuring that all discussions and analyses remain part of the review record. Neo operates with the same RBAC permissions as the user's Pulumi Cloud account, maintaining the same level of security and governance. These integrations are designed to streamline workflows by allowing engineers to engage with Pulumi Neo from the communication tools they already use, without granting additional permissions. The launch is part of Pulumi's broader vision for agentic infrastructure, enhancing accessibility and usability for teams with linked Pulumi Cloud identities.

Pulumi Neo, initially available through Pulumi Cloud, now offers a seamless terminal experience, enabling users to execute tasks locally without the need for additional setup or credentials, as it inherits existing configurations such as authenticated CLIs and environment variables. This local execution is ideal for interactive sessions where users and Neo collaboratively tackle infrastructure tasks, while Pulumi Cloud Neo remains suitable for autonomous tasks that require asynchronous execution. Users can also delegate tasks to Neo from other agents, leveraging its integration with tools like Atlassian, Datadog, and PagerDuty. The consistent identity, RBAC, and audit capabilities between terminal and cloud ensure secure operations, and users can start using Pulumi Neo by authenticating through Pulumi Cloud, with further guidance available in the Pulumi Neo documentation.

Pulumi Neo introduces new capabilities to enhance infrastructure management by integrating with third-party systems and cloud command-line interfaces (CLIs), allowing seamless execution of tasks across various platforms. The integrations include connections to services like Atlassian, Datadog, Honeycomb, Linear, PagerDuty, and Supabase, as well as CLI access to AWS, Google Cloud, Azure, and Kubernetes, all configured at the organization level. An illustrative scenario demonstrates how Neo can autonomously manage infrastructure alerts by coordinating data from PagerDuty and Datadog, adjusting settings via AWS CLI, and orchestrating changes through Pulumi, all within a single conversation. These features aim to streamline infrastructure investigations, provide live cloud insights, and ensure per-task control and failure handling, enhancing the efficiency and reliability of infrastructure management. The launch signifies a step toward agentic infrastructure, inviting user feedback and suggestions for further integration improvements.

Pulumi has evolved its infrastructure-as-code platform to embrace the emerging era of agentic infrastructure, where AI agents play a significant role in coding and managing IT infrastructure. This transformation leverages Large Language Models (LLMs), which have proven adept at coding, to automate and enhance infrastructure tasks by mapping infrastructure space to code space. Pulumi supports numerous programming languages, allowing for the application of software engineering patterns to infrastructure management, and has developed tools to make infrastructure verifiable and auditable. By integrating with AI models like Anthropic’s Mythos and OpenAI’s Codex, Pulumi aims to increase the autonomy of agents in handling infrastructure, while maintaining crucial human oversight through robust change management and policy enforcement systems. Recent platform updates include agent-friendly features, new integrations, and a benchmark for measuring agent performance on infrastructure tasks, as Pulumi continues to push towards a fully agentic future in infrastructure management.

Pulumi has released an updated version of its Command Line Interface (CLI) designed to enhance usability for both human developers and AI agents, emphasizing guessability, accessibility, and readability. The new CLI structure simplifies command guessing by using singular nouns for branches and verbs from a canonical vocabulary, making it easier to navigate tasks without needing detailed explanations. This reorganization allows users to execute all Pulumi Cloud operations directly from the terminal, providing consistent command naming and output formats. The CLI now supports direct API access without managing separate tokens, streamlining interactions with the Pulumi Cloud. Users can also discover templates directly from the shell and access agent-friendly Markdown documentation for packages and components, reducing verbosity and enhancing efficiency. Additionally, Pulumi Neo, which was previously only available through the Pulumi Cloud Console, is now integrated into the CLI, offering developers a complete terminal-based experience. The release also includes improvements in JSON output handling, consistent exit codes, and enhanced help text, with a sneak peek at a new command, pulumi do, for direct resource operations. Pulumi encourages feedback to continue refining the CLI for both human and agent users.

Pulumi Neo has evolved significantly since its launch, enhancing its capabilities to support platform teams in automating and streamlining infrastructure tasks. Initially focused on simplifying cloud deployment, Neo now integrates with the Pulumi CLI, GitHub, and Slack, offering diverse functionalities such as deploying applications to AWS, diagnosing API performance issues, triaging alerts, and implementing tickets from tools like Linear and Jira. Neo also aids in auditing and tightening IAM roles, migrating services to Kubernetes, and managing infrastructure drift with scheduled checks. By automating routine tasks such as weekly upgrades for outdated runtimes and fixing compliance issues like CIS Benchmark failures, Neo acts as a virtual platform engineer, leveraging organizational context and established conventions to propose changes and improvements via pull requests. This approach shifts the role of platform engineers from hands-on management to oversight, enabling them to delegate tasks and focus on strategic decision-making.

In May 2026, Pulumi announced several advancements aimed at enhancing agentic infrastructure, where both AI-driven agents and humans manage infrastructure tasks. Key updates include improved command-line interfaces (CLIs) to facilitate better interaction for coding agents and humans, with new functionalities such as one-command execution and imperative operations across Pulumi-supported clouds. Pulumi introduced ephemeral agent accounts for seamless infrastructure management and expanded Neo, its infrastructure agent, beyond the Pulumi Cloud console into CLIs, Slack, and GitHub, enabling automation and integration with services like Atlassian and Datadog. Collaborations with AI infrastructure leaders led to new partner providers, including CoreWeave and NVIDIA AI Cluster Runtime, to support AI workloads with advanced GPU infrastructure. Additionally, Pulumi has updated its documentation to be more agent-friendly and introduced InfraBench, a benchmark to assess agent performance on infrastructure tasks, reflecting a shift towards more intelligent and automated infrastructure management.

In recent years, the process of building AI agents has shifted significantly from requiring extensive infrastructure setup to a more streamlined approach due to advancements in software development kits (SDKs) and integrated tools. Previously, developers had to manually set up complex retrieval-augmented generation (RAG) pipelines and write significant amounts of custom code to enable agents to perform their tasks. However, the introduction of built-in tools like the Claude Agent SDK and OpenAI’s Codex SDK has simplified the process by providing essential functionalities such as file handling, shell commands, and web interactions out of the box. This change has reduced the need for middle-layer infrastructure, allowing developers to focus more on the agent's core capabilities rather than on setting up the environment. The shift has also led to the adoption of a skills-based approach, where agents load tools only when needed, resulting in more efficient use of resources. While traditional frameworks are still relevant for specific use cases such as multi-agent orchestration or deterministic typing, starting with an SDK is generally recommended for most projects, as it often covers all necessary functionalities without the overhead of a full framework. This transformation aligns with existing infrastructure practices, emphasizing integrated tools and governed actions, and allows developers to build more efficient and focused AI agents.

The concept of a "dark factory," originating from Fanuc's robotics plant in Japan where robots operate without human presence, is now emerging in software development, where autonomous systems handle code generation and validation with minimal human intervention. This trend is exemplified by companies like StrongDM and Stripe, which have implemented varying levels of software autonomy as outlined in Dan Shapiro's autonomy ladder—ranging from basic automation to fully autonomous systems that require no human oversight before code deployment. Infrastructure presents a more complex challenge than application code due to factors like blast radius and state management, but tools like Pulumi offer components that facilitate a more controlled transition to autonomous operations, emphasizing the importance of isolation between code generation and validation to prevent sycophantic behavior from models. The key to successful implementation lies in establishing "holdout scenarios" for isolated validation and gradually expanding the scope of automation as confidence in the system grows, while maintaining tight control over destructive operations and regular audits to mitigate risks.

Custom VCS is a new feature in Pulumi Cloud that enables integration of any Git or Mercurial version control system with Pulumi Deployments through webhooks and centrally managed credentials, addressing the limitations faced by teams using self-hosted or third-party VCS systems. Previously, these teams had to manually configure credentials for each stack without automated deployment triggers, but Custom VCS introduces an organization-level integration that allows for centralized credential storage in a Pulumi ESC environment and webhook-driven deployments. This setup enables multiple stacks to share credentials, supports push-to-deploy functionality, and allows for manual repository registration, though it lacks some features available with native integrations like pull request comments and commit status checks. Pulumi's AI assistant, Neo, can perform repository operations under Custom VCS, although it cannot handle tasks requiring VCS-specific APIs, such as opening pull requests. To implement a Custom VCS integration, users need to configure their VCS server to send webhooks to Pulumi and manage credentials through the ESC environment, with more detailed guidance available in the Custom VCS documentation.