Harness bought Augment’s coding agents to show them what happens after the merge
Harness announced Thursday that it acquired Augment Code’s Cosmos software factory, Auggie CLI and Code Context Engine, along with the team behind them, for an undisclosed amount. The deal adds autonomous coding agents to Harness’s existing software delivery platform, which already handles testing and deployment.
According to Harness, Cosmos will become the Harness Cosmos Software Factory Agent, and the company says the connection between Augment’s codebase understanding and Harness’s delivery history is already live. Cosmos, the Code Context Engine and the Software Delivery Knowledge Graph are linked to Harness through the Model Context Protocol (MCP), so teams writing code in Cosmos today have context from across the software development lifecycle available to the agent as it works.
“AI’s impact on software will not be measured by how much code it generates, but by how much valuable software reaches customers,” said Harness founder and CEO Jyoti Bansal. The acquisition follows the company’s research into the AI velocity paradox, which found that faster code generation was putting additional strain on testing and release processes that had not kept pace.
“AI’s impact on software will not be measured by how much code it generates, but by how much valuable software reaches customers.”
Inside the Cosmos factory
Cosmos takes a requirement or bug report through to a pull request, with agents writing and testing code inside isolated VMs. They can also pick up failed checks and review comments on an existing PR, using feedback from earlier rounds to make further changes.
Teams can fork and customize agents for their own workflows, while Augment’s Code Context Engine retrieves relevant codebase context as they work.
“We set out to bring AI to the entire software development lifecycle, starting with context-aware coding,” said Igor Ostrovsky, Augment Code’s co-founder and CTO, describing engineering work that becomes increasingly autonomous “with engineers stepping in only where judgment really matters.”
“We set out to bring AI to the entire software development lifecycle, starting with context-aware coding.”
Where humans still approve
Developers still approve and merge changes, although Bansal wrote that teams can decide which work agents handle on their own and where a person has to review or approve it.
Cosmos is still reached through Augment’s existing site for now, and Sanjay Nagaraj, senior vice president of global engineering at Harness, tellsThe New Stack that access will move into the Harness platform in the future, though the agent already has Harness’s deployment history and security findings in hand no matter where developers open it.
Turning failures into fixes
Harness described to The New Stack the division of labor between Augment’s Code Context Engine and the Software Delivery Knowledge Graph in terms of what each system knows, with code context telling the agent what the code does and the delivery graph telling it what happens once that code ships, so the agent can write changes that are safe to deploy rather than ones that are only correct in isolation.
“When Harness downstream agents identify an issue while testing, securing, or deploying code, they send it back to Cosmos with the context to fix it,” Bansal said. Nagaraj confirmed that this return loop is already running, though not yet in its final form.
“When Harness downstream agents identify an issue while testing, securing, or deploying code, they send it back to Cosmos with the context to fix it.”
“Today it runs through the ticket workflow,” he explained to The New Stack, adding that when a downstream agent flags a testing, security or deployment issue, the relevant context is packaged into the ticket for Cosmos to act on, with “a more native workflow” coming soon.
A failed deployment therefore already lands back with the coding agent carrying its failure context instead of leaving a developer to diagnose the issue and open a separate task, even if a ticket still sits in the middle of that handoff for now.
How the agent decides which deployment records or security findings to pull for a given change matters for teams weighing how much production history an agent should see before it touches their code.
Nagaraj tells The New Stack, “standard role-based access controls apply,” and that within those boundaries Cosmos selects context using the code graph and the relationships it maps between a change and the services, deployments and findings connected to it.
Who controls agent permissions
Harness plans to bring Cosmos under the identity and policy controls of its Agent Harness, work that Nagaraj said “is in progress,” and the company hasn’t explained how permissions will work across the two systems once that lands, a gap that stands out as the industry moves toward agent identity, with Microsoft giving its Copilot agents identities of their own.
Cosmos runs agents in isolated VMs with configurable review checkpoints, and while role-based access controls set the outer boundary on what production data the agent can reach through the delivery graph, it remains unclear which downstream actions it can take once a failure is handed back to it. Those limits deserve scrutiny, particularly after coding agents exposed more than 13,000 screenshots through routine workflows without an attacker ever gaining access.
Update, Oct. 14, 2026: This story has been updated with comments from Harness SVP of Global Engineering Sanjay Nagaraj confirming that the failure return loop is live through a ticket-based workflow, that Agent Harness identity and policy integration is still in progress, and how Cosmos selects context from the Knowledge Graph.
