Pulumi Unveils Delivery and Agents in Public Preview

Pulumi introduced Pulumi Delivery and Pulumi Agents on Wednesday, now available in public preview, to further automate software publishing and remediation. There are limits on the modifications allowed, while some features require a Pro or Enterprise subscription.
Customers Choose Their Agents at Pulumi
Launched last year to work on infrastructure, Neo is now just one possible agent in Pulumi's offering. Customers can configure as many as they wish, assigning each a mission, guidelines, and tools, or host their own agent, such as Claude Code or Codex, in the company's cloud. Primarily based on Claude from Anthropic until now, Neo can also utilize models from OpenAI or Google.
A request, a scheduled job, or an alert from PagerDuty, Datadog, or Honeycomb can trigger their intervention. Reports are sent to teams via Slack. Many customers are already using their Claude Code sessions to utilize Pulumi, thus settling invoices with both Anthropic and Pulumi. Despite the additional consumption associated with the agents, gross margins remain in the high 80% to low 90% range, claims Joe Duffy, co-founder and CEO of Pulumi.
The company has showcased these products to around 50 customers; a dozen are participating in a private beta. Duffy mentions a reception that exceeds all previous launches and cites Moderna among the companies seeking access. Pulumi has thousands of customers, including BMW, Snowflake, and Wiz.
Teams Set Acceptable Risks
With Delivery, engineers and agents describe the path of a release in code: testing, validation, and then deployment. An AI review system can assess the risk of a modification before it is published or request human intervention. Pulumi Agents, on the other hand, focuses on failure review, cloud cost savings, and security issue resolution. Each agent has a limited modification scope, and their interventions are presented as code proposals that the team can review.
Delivery can leverage Jev, the TypeSafe AI model, to quantify the probability of specific risks that the team wishes to evaluate before a modification. Exceeding a threshold set by the team blocks the publication. Daniel Gafni, a technical staff member at TypeSafe AI, explains that this choice of whether to publish or not corresponds to the type of decision aimed for from the design of Jev. Access to Jev's decision controls and agent configuration is part of the features reserved for Pro or Enterprise subscriptions.
Duffy recounts discovering TypeSafe in May when it became a customer of Pulumi: its name had caught his attention. A few months later, with the rise of Jev, TypeSafe used Pulumi to deploy globally.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Pulumi Aims for Post-Merge Work
According to Duffy, writing, reviewing, and merging code constitutes the limit of many assisted programming tools. The subsequent steps, from testing to publishing, are handled by Pulumi's new products. The executive suggests that these products could take over certain tasks currently assigned to Argo CD, GitHub Actions, or internally developed scripts.
Duffy cites Harness, which uses the term "agentic SDLC" to refer to the software development lifecycle. GitHub showcased agentic software delivery at its Universe conference last month; Cursor also aims to tackle this issue, according to Duffy, while Anthropic has an internal system. The CEO describes a major transformation in the industry and believes his company is particularly well-positioned to respond.
Founded in 2017 by Duffy and Eric Rudder, both former Microsoft employees, Pulumi initially offered to manage servers, networks, and permissions using languages like Python or TypeScript. This choice suits agents trained to produce code in these languages, an advantage Duffy admits he did not foresee. According to him, agents now perform 40% of the operations previously done by humans on the platform, up from 15% in April and none in January. Claude Code, Codex, and Neo thus autonomously modify the infrastructure.
Pulumi Develops Its Products with Fewer Employees
Pulumi has around 100 employees, down from about 130 last year. Duffy states that no layoffs have occurred and attributes the evolution of roles and increased efficiency to AI; he claims that annual recurring revenue per employee has more than doubled over the past year. About 40 employees work in the Seattle area, at the Two Union Square office. The company is hiring more there, as the executive believes physical presence is important for developing new products, and is also looking for salespeople in nearly every region of the world.
Duffy reports good results with product-focused engineers who can adopt the customer’s perspective and lead agents like a team. He credits AI with enabling the realization of projects long considered too ambitious. The fiscal year that began in July has thus been dubbed "the year of absurd ambition": in a message to employees, the executive presented ambition, rather than speed of development, as the limit to their achievements.
The company is not yet profitable, despite a few months of positive cash flow this year. Its latest funding round, led by Madrona in 2023, brought in $41 million, for a total of approximately $99 million raised. Duffy states that investors initiated the last two rounds while Pulumi was not seeking capital, and that each deal was completed in about 48 hours.
According to the executive, Tim Porter from Madrona has joined the board in place of the late Soma Somasegar, who led the first investment in Pulumi. When asked about a potential sale or another fundraising round, Duffy mentions several paths to build a successful long-term company, while expressing a desire to choose the conditions.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.