▶ Watch ↗AI Engineer World's Fair 202514:53
Conference affiliation: Principal Product Manager · Microsoft · 2026
Harald Kirschner is a principal product manager at Microsoft working on Visual Studio Code and GitHub Copilot. His work spans the mechanisms that let coding assistants retrieve project context, connect to external services, and carry plans across sessions. He brings experience in browser libraries, Firefox performance measurement, and developer tools to the design of agent workflows that people can inspect and direct.
Harald’s early career combined JavaScript consulting with open-source development. Working under the alias digitarald, he contributed to the MooTools core team and published browser libraries. His FancyUpload replaced a basic file input with multiple-file selection, queued uploads, and animated progress indicators, making the progress of several uploads visible to the person waiting for them.
From 2009 to 2011, Harald led web front-end development and partner integrations at Poken in Lausanne. He joined Mozilla in October 2011 as a partner-engineering manager, helping applications and games ship on Firefox OS and supporting outreach around browser gaming. In 2016, he moved into product intelligence, coordinating user-centered performance metrics for Firefox Quantum and leading the initial development of the health.graphics KPI dashboard. From 2017 to 2020, he was staff product manager for Firefox DevTools, responsible for developer-experience strategy across teams. His career history includes extending that work into partner integrations, testing, automation, and performance tooling.
His writing on Firefox’s JavaScript debugger explains the team’s improvements through specific developer problems: lost breakpoints, pauses in the wrong script, slow variable previews, and source maps that exposed compiled output instead of the original code. The debugger team and its contributors implemented the changes; Harald explained their purpose and gathered feedback about remaining gaps. He also called for explicitly reserving time for polish and continuing to test against changing frameworks and compiled languages.
In his Webhint article, Harald described collaboration between Firefox DevTools and the Webhint team to bring compatibility, accessibility, security, and performance guidance into page inspection. The point was to find problems while developers could still act on them, rather than discovering a compatibility failure late enough to require a rewrite. The hint engine was extensible, allowing teams to encode their own practices rather than treating every recommendation as a universal rule.
Harald joined Microsoft as a senior product manager in October 2020 and became a principal product manager in August 2023. His work on VS Code’s developer experience and growth expanded into GitHub Copilot and the problem of answering questions inside an existing repository.
In his identifiable exchanges and workspace demonstration in 2024, Harald explained how Copilot combined local sparse keyword retrieval with GitHub-hosted semantic indexes and language-model reranking. Local retrieval let workspace questions work even when a repository lacked a hosted semantic index. Code-aware segmentation preserved functions and blocks, giving the assistant meaningful program structure when answering questions such as how benchmarks run or which tests a repository contains. These were components of the team’s retrieval system, rather than one search method applied to every project.
Harald also described evaluation as an early investment: smaller regression suites could run on pull requests, while larger daily suites covered repositories across languages. Executable code tests checked behavior. This gave the team a way to test whether retrieval and prompting changes worked beyond the examples that happened to look convincing during development.
Harald authored VS Code’s May 2025 account of MCP integration. It describes a consequential design transition: the team initially used VS Code’s Tools and Chat participants APIs, then adopted Model Context Protocol so services could connect across clients through a shared standard. Installation commands, securely stored input variables, server logs, and a per-session tool picker made those connections usable within the editor.
In June, Harald coauthored the full MCP specification support announcement with Connor Peet and Tyler Leonhardt. His full-spec MCP talk develops the argument that tools alone leave much of the protocol’s value unused:
He also acknowledges the ecosystem’s tools-first starting point: VS Code initially followed that approach too. His recommendation is progressive enhancement—implement the richer capabilities, test them in supporting clients, and use developer feedback to improve the specification and SDKs.
Harald’s approach to enterprise coding agents leaves room for disposable, outcome-first experimentation. A working prototype can help a designer communicate an idea or give someone an unfamiliar codebase to learn from. He treats that freedom as appropriate to exploration; sustained team development needs more explicit decisions about the stack, conventions, and implementation.
His structured approach starts with a template containing the organization’s technology choices, design system, project instructions, and tools for relevant knowledge. That reduces the chance of handing engineering an attractive prototype built around conventions the team does not use. Instructions guide how the agent works; reusable prompts capture tasks the team wants to perform consistently. For larger changes, spec-driven development separates the specification, implementation plan, and execution.
The review points are concrete. In a test-driven custom agent, Harald has the assistant write tests first and pause for his confirmation before implementing, because reviewing the tests is easier for him. Linters and task runners expose errors the agent can act on; frequent commits and checkpoints preserve working states. He also limits auto-approval to a workspace when demonstrating it and notes that command chaining complicates terminal allowlists. Selecting fewer tools narrows the agent’s options, but he explicitly acknowledges that instructing a model to use a tool does not guarantee it will do so.
Harald’s Agent TODOs project adds persistent task records, tools for reading and updating them, and an Explorer view of pending, active, and completed work. A plan can survive a session restart instead of depending entirely on what remains in the conversation.
He contributed alongside other VS Code team members to the unified coding-agent experience, covering agent-session management, planning, delegation, and separate subagent contexts. In February 2026, Harald described tuning the Plan agent for new models and orchestration performance. He also demonstrated Agent Skills integrating with extension APIs, connecting agent workflows to the editor’s existing extension ecosystem.
His MCP Apps Playground explores another way to keep people involved: interactive interfaces inside agent conversations. A flame graph lets a developer zoom into a performance profile, investigate a suspected hot path, and send a selected frame back to chat. A list-sorting interface lets the user change a proposed order directly, supplying preferences that would otherwise require repeated messages. These examples extend his developer-tool work into agent interactions: the person can examine the underlying data and revise the result, rather than relying on the agent’s summary alone.
▶ Watch ↗AI Engineer World's Fair 202514:53
▶ Watch ↗AI Engineer World's Fair 20241:19:45
▶ Watch ↗AI Engineer World's Fair 20251:19:33
▶ Watch ↗AI Engineer World's Fair 20241:29:02
▶ Watch ↗AI Engineer World's Fair 20251:20:38
▶ Watch ↗AI Engineer World's Fair 202515:26
Loading more talks and highlights…