Augment Code launches Cosmos to bring agentic AI software development to teams

Augment Code Computing Inc., an artificial intelligence agent platform provider, Thursday announced the launch of Cosmos, a service it says is designed to push beyond the era of individual AI coding assistants and into a more coordinated model for engineering teams.

In an interview with SiliconANGLE, Vice President of Engineering Vinay Perneti said that as AI has progressed, users and software engineers alike have moved along with it. It’s time that the software industry received the benefit of the technology.

“I think the eras were almost like ’24 was probably mostly chat, ’25 I think is agents, but I would actually go a little step further…. ’26 is going to be agents for teams,” he explained.

Perneti says Cosmos represents the next phase of that cycle, where agents have been inserted into the software development lifecycle doing autonomous work for individual developers and seeing their speed increase. But teams haven’t seen that same boost because context isn’t shared between them.

Agentic AI has promised that a developer can hand off work to a virtual worker that knows their codebase. It understands their personal style, can “see” what they’re working on and take the most tedious, most annoying and worst parts of coding off their hands — or at least that’s what’s on the brochure.

“Every single leader, engineering leader, that I talked to was doing top-down, you have to adopt agents… but then they were all equally unsatisfied when they were looking at the team throughput,” Perneti said.

Some people on teams using agents will see huge spikes in throughput because suddenly they can offload labor to agents. However, as a whole team, this experience is better described as “lumpy” because not everything a team does can be described as easily augmented and the context of what one person does isn’t visible across the entire suite of agents all the time.

Perneti said the company developed Cosmos to solve this problem.

Cosmos operates at a higher level than just an agent per developer. The system still provides individual developers agents so that they can spin up an agent to help them code, but it also delivers entire internal organizations the ability to plan, develop, build and deploy using cross-team knowledge. This is the bottleneck the platform is tackling.

The company calls it a software delivery lifecycle platform, and it’s designed to give humans and agents a way to collaborate. Agents handle the bulk of execution and work autonomously, but humans remain involved in prioritization, specification, intent validation and high-risk decision-making.

“Agents are still not good at making good long-term decisions… so humans are still very much in the loop on making sure the spec is very high-quality,” Perneti said.

For example, developers would still be thinking out loud about what they want as if directing a junior coder. They’d provide a specification document alongside what they want, the key elements and the primary goal. This becomes the guidepost for the agents to gather around and work against. To make sure that this makes sense, a validator agent can also look it over and come back to the developer and clarify anything that needs to be updated or verified against the codebase and company best practices that have been built up over previous work.

Perneti said Cosmos was designed to build this knowledge not just from looking at code, but to learn over time from every touchpoint. When working on a project, a Slack agent might receive a correction from one coworker, will carry forward by keeping a memory scratchpad, maintaining a shared conception of patterns, corrections and other practices represented in trends between sessions, agents and teammates.

Cosmos is more than just a coding tool

Because Cosmos is cross-team, it allows non-experts to access knowledge and data that would otherwise be sectioned off or moderated through bureaucracy or technical know-how.

For example, the sales team does not need to set up its own Claude Code, BigQuery credentials, query tooling or talk to the analytics team just to get access to business intelligence. Now salespeople can go into Cosmos and get access to an agent set up by the data team.

“Our data team set up one expert, called a data analytics expert,” Perneti said. “It’s a read-only expert that has all the credentials configured, so the sales team can just go… ask this expert.”

This data analysis agent is still fully under the control of the data team. Everything it knows maintains full, real-time access to business knowledge, optimization and governance – but it doesn’t suffer from drift or bottlenecks of having to get approvals or the organizational stress of being fractured across the organization. When the data team updates, it also updates for the sales team.

“Cosmos is what I guess I call self-aware, meaning we made Cosmos aware of all the other sessions and all the other agents that are running,” Perneti added.

Building on the cross-team functionality, Cosmos includes an adviser that can help users pick the right expert and environment for a task. This matters in an organization where there might be dozens of experts and agentic workflows that could get any particular task completed. Instead of requiring users to know which particular agent to invoke, they can go to Cosmos and simply describe what they need and the adviser will take over, route the work to the correct agent, and get it underway.

Image: Augment Code