Jiachen Jiang

Jiachen Jiang is a Staff Product Manager for AI Governance at Coder. Their work addresses a practical tension in coding agents: they need access to repositories, services, and credentials to do useful work, but that access also lets a mistaken or compromised agent cause damage. Jiachen explains how execution environments, permissions, and independent logs can constrain those consequences. Their earlier product work at Microsoft covered NuGet, .NET, and Azure Container Apps, connecting developer workflows with the infrastructure behind them.

Developer feedback and deployment workflows

Jiachen earned a B.A. in Computer Science from Dartmouth College. At Microsoft, they worked first on the .NET team and later on Azure Container Apps; their .NET product work included NuGet, Entity Framework, and ASP.NET.

Their 2021 examination of the NuGet ecosystem brought together survey and interview findings about package discovery, package authoring, and everyday workflow friction. A subsequent NuGet.org package-page redesign showed how user feedback could change a specific interface decision. Embedded READMEs and planned compatibility information needed more room, so the team introduced tabs. But a proposed dropdown for installation instructions concealed choices users wanted to compare and required an extra click. The team restored the previous instructions after feedback. Jiachen documented that reversal and credited summer intern Sophia with driving the page’s design and implementation.

On Azure Container Apps, Jiachen wrote about deploying distributed applications without requiring every developer to manage a Kubernetes cluster. Their deployment guide for .NET Aspire connected local development of interdependent services with cloud provisioning and deployment through the Azure Developer CLI. It addressed the work between running an application locally and operating it in the cloud.

Constraining what coding agents can do

Jiachen’s writing on agent security describes agents as needing flexible access like people while warranting the limited trust given to machine workloads. That combination calls for controls enforced by the environment: a written instruction cannot revoke a credential or prevent a network request.

In “Your AI Isn’t the Problem. Your Infrastructure Is.”, Jiachen uses the public Replit database incident to make the distinction concrete. A code-freeze instruction did not remove the agent’s ability to delete production data. Logs within an agent’s reach also cannot provide an independent account of its actions. Jiachen argues for separating agent execution from production access, limiting credential scope and lifetime, and keeping audit records outside the agent’s control. These layers limit different failure paths; none guarantees that an authorized action is correct.

Their Coder writing develops several related infrastructure decisions:

• Process-level network controls: Jiachen explains why an IP allowlist can be too coarse for an agent’s work. Coder’s Agent Boundaries, subsequently renamed Agent Firewall, supports restrictions by destination and HTTP method. An agent can retain access to a repository or package registry while being blocked from sending data to an unauthorized destination. Jiachen articulates the mechanisms and tradeoffs of these Coder team products; the work is not presented as their sole invention.

• Shared environments with individual access: Their account of Shared Workspaces addresses collaboration across files, services, running processes, and persistent state. Individual authentication lets colleagues enter the same environment without exchanging personal credentials. Service-account ownership lets an operational workspace persist beyond one person’s involvement—for example, an on-call environment that passes between shifts.

• Measuring delivery rather than code volume: In their approach to governed agent deployment, Jiachen combines isolated workspaces, visibility into model traffic, and process-level enforcement with measures of engineering outcomes. They recommend tracking cycle time, defects after merge, and cost per merged pull request. These measures ask whether agents improve delivery and at what cost, rather than treating more generated code as success.

Jiachen’s security approach also distinguishes permission from correctness. Keeping an assistant within its user’s permissions limits what it can reach, but the assistant can still make a damaging mistake within those permissions. Their practical recommendation is to locate persistent tokens, routes to production, and unsandboxed execution, then rehearse an incident before those weaknesses are tested by a real failure.