From Solo Helpers to Team Infrastructure
AI coding agents for teams are shared, orchestrated systems that coordinate multiple agents, workflows, and approvals across an engineering organization instead of acting as isolated assistants for individual developers. In the first week of June, Cognition, Microsoft, and Augment Code all launched platforms that move AI coding agents into this shared layer. Cognition’s Devin Desktop, Microsoft’s Rayfin, and Augment Code’s Cosmos sit at different points in an emerging stack that treats agents more like CI pipelines than autocomplete tools. Rather than one developer running a local agent loop, teams gain a “harness” that remembers decisions, enforces policies, and coordinates many agents in parallel. As one analysis notes, the contest is shifting from model quality toward the orchestration, review gates, and control planes that decide what agents run, when, and how their changes reach production.
Dropbox Nova Shows Enterprise-Scale Orchestration
Dropbox’s Nova platform is an example of AI coding agents teams infrastructure built directly into engineering workflow automation. Nova does not present a single assistant; it is a centralized execution layer where AI agents operate inside the company’s Bazel-based monorepo, CI pipelines, observability stack, and infrastructure tools. Each Nova session runs in an isolated cloud environment tied to a specific commit, following a “propose, validate, iterate” loop against real builds and tests before any change is accepted. This shifts agents from local code generators to actors in production-grade workflows such as flaky test remediation, dependency migrations, and incident investigation. According to Dropbox, the platform was created to close the gap between off-the-shelf coding tools and the realities of large-scale, highly customized engineering systems, turning collaborative coding platforms into core operational infrastructure.
Cognition’s Devin Desktop: A Control Console for Agent Fleets
Cognition’s Devin Desktop brings the team infrastructure orchestration pattern directly to the developer desktop. The product combines a full code editor with an agent coordination dashboard, letting teams manage local and cloud agents across projects, repositories, and environments. It keeps the Windsurf editing experience, but adds an agent management layer around it instead of forcing teams to rebuild workflows. Spaces group agents by project and share context across sessions, pull requests, files, and tasks so work is organized around real engineering units rather than one-off chats. Theodor Marcu, Head of Product Growth at Cognition, said that leaders now must manage “a growing fleet of agents working across their organisation simultaneously.” Support for the Agent Client Protocol (ACP) opens Devin Desktop to third-party and internal agents, turning it into a vendor-neutral console for collaborative coding platforms rather than a single-agent product.
Sandboxes, Security, and Infrastructure for Agent Workloads
As AI coding agents move deeper into shared systems, agent sandbox security and execution infrastructure become central concerns. OpenAI’s work on a secure Windows sandbox for Codex shows how hard it is to run autonomous agents on developer machines without risking the host environment. Codex needs direct access to tools and repositories, yet it must not have unrestricted system access. OpenAI evaluated built-in Windows options such as Windows Sandbox and Mandatory Integrity Control, then built a custom “unelevated sandbox” combining security identifiers (SIDs), access control lists (ACLs), and write-restricted tokens. This mirrors a broader trend: infrastructure providers now design sandboxed environments specifically for agent tool use and model evaluation, while team platforms like Nova and Cosmos run agents in isolated sessions tied to commits and policies. The result is an architecture where orchestration, isolation, and auditability are first-class features of AI-powered engineering workflow automation.
