⚡ Inside GitHub’s AI Revolution: Jared Palmer Reveals Agent HQ & The Future of Coding Agents

Summary

• GitHub is positioning Agent HQ as a hoped-for home for coding agents and developers, not merely another first-party assistant. The host cites 183 million developers, while Jared Palmer says GitHub has over 180 million. Jared says GitHub can pair Copilot and Copilot CLI with third-party harnesses including Claude Code, Codex, and Devin, creating a “gravity well” for agents and developers. The strategic asset is workflow ownership plus agent choice, rather than dependence on one model.
• V0’s history suggests model progress monetized explosively only after product form and scope caught up. Jared says the initial product took roughly nine months to reach $1 million ARR; the host says that after models could support chat, a rewrite added “another million” roughly every 14 days and became “a rocket ship.” Jared credits the deliberately narrow focus on Next.js, frontend, and shadcn: “That constraint was rather liberating.”
• The relevant abstraction is shifting from a model switcher to a tightly coupled agent with compute, files, tools, and a runtime. The discussion argues that loose model interfaces collapse toward the lowest common denominator. Jared calls the coding-agent setup a distinct “agent world,” with sandboxing, file systems, and tool calls. Value therefore shifts toward harnesses, specialist subsystems, evaluations, and infrastructure around the model.
• GitHub’s near-term wedge is ambient AI across the software lifecycle, with VS Code integration as leverage. The demonstrated flow sends an Agent HQ task into a pull request and opens it in VS Code with one click; Jared also points to issue assignment, merge conflicts, and broken Actions as intervention points. His focus for roughly the next six months is AI sprinkled “salt-bae style” through mobile, web, GitHub.com, and editors.
• Reliability, not another flashy modality, is the remaining slope from 90% success to “nines of success.” Jared’s answer to what comes after today’s coding agents is simply “making them good”: the gap between 98% and 99% correctness is noticeable, while multi-turn sessions amplify provider failures. At Vercel, key metrics rolled up about every three hours, including error-free sessions, latency, dropped requests, and infrastructure errors.
• Reproducible repository setup remains an unsolved infrastructure bottleneck, and dev containers are promising without yet being the standard. A focused Next.js system can assume a package-install path; a general agent cannot predict whether a repository quietly needs something like ffmpeg. The discussion notes competing runtime approaches inside Microsoft and points to shared defaults, autodetection, and preemption as open opportunities.
• GitHub is pairing Agent HQ with overdue core-product repairs, but its scale makes even popular requests expensive to land. A redesigned homepage now emphasizes tasks and recent pull requests, while stacked diffs—the community’s top feature request—are back under active exploration. Internal attempts date to 2020, including a polished 2022 implementation deemed too risky because it introduced stacks outside pull requests.

Deep dive

1. V0’s narrow constraint converted model progress into a product

• Jared’s origin story predates any Vercel AI group: while directing engineering across Next.js, Turborepo, Turbopack, webpack, and internal tooling, he turned a Server Actions dogfooding exercise into an AI playground. Guillermo’s news that nat.dev would launch Monday forced him to ship; the exercise made him work through every provider’s API and streaming quirks.

• That playground became AI SDK because Jared had already isolated the reusable streaming layer without forcing developers into one provider. Its niche was the UI layer: useful abstractions without getting in the developer’s way. An open-source ChatGPT template built with shadcn then gave the small team a full-featured base for rapid experimentation: “It was liberating.”

• Code Interpreter sparked the original generative-UI thesis: code execution could emit tabular data, another prompt could render it, and outputs might pipe together. Yet tool calls did not exist, the context window was 4,000 tokens—later about 16,000—and internet-enabled execution triggered a security debate. The code-execution idea was shelved; prompt-to-UI became v0’s “aha moment.”

• Around September 2023, v0 launched as “Midjourney for React,” using selectable visual variants because reliable chat still was not available. Progress through GPT-4 and GPT-4 32K changed the ceiling, though Jared admits, “We never really got GPT-4 Turbo working—I don’t know why.”

• Commercially, Jared says the first version took roughly nine months to reach $1 million ARR. The host then says that once models could sustain chat and the artifact pattern matured, the team rewrote the product; the relaunch added “another million” roughly every 14 days and became “a rocket ship.” The durable edge was focus on Next.js, frontend, and shadcn. The host also says Vercel shared its post-training harness and hygienic data with frontier labs.

2. The agent is replacing the standalone model as the product unit

• The conversation describes a debate between combining the best qualities of different models and exposing a model selector. A composite permits specialist subsystems—search need not use the generation model—can improve performance by stitching components together, and gives the product a brand it controls.

• The economics cut both ways. A branded provider model can earn co-launch promotion, but billing is effectively anchored near that provider’s retail price, limiting the premium. Jared adds that a composite can support a sustainable business independent of model labs while genuinely improving performance.

• The hosts argue that the model layer is no longer the right switching abstraction: models and agent harnesses need to be “very, very strongly coupled,” or a generic interface inherits the lowest common denominator. Jared characterizes the coding-agent setup as a distinct “agent world” involving loops, compute runtimes, files, sandboxing, and tool calls.

• Agent HQ reflects that shift. GitHub supplies Copilot and Copilot CLI while also accommodating third-party harnesses such as Claude Code, Codex, and Devin. Jared hopes it becomes “the home” of agents and developers, offering choice rather than forcing one first-party stack.

• Skills push the interface down another layer: the host frames them as an LLM-native bundle where an agent reads Markdown and a directory of files, making the filesystem a universal interface. The discussion highlights MCP as an important enterprise-context mechanism, while GitHub’s announced custom agents can combine task-specific prompts with MCP capabilities.

3. GitHub’s advantage is owning the seams between development tools

• V0 optimized one framework and one problem; GitHub spans all languages, frameworks, and the 183 million developers cited by the host. Jared, only on day 13, calls that scale “immense” and places Agent HQ in a broader Microsoft CoreAI organization containing GitHub, Visual Studio, VS Code, and parts of Azure.

• The intended experience is continuous rather than destination-based: launch a task, receive a pull request, then open it in VS Code with one click. Jared wants AI inserted at existing friction points—assigning issues, resolving merge conflicts, or repairing an Action that fails remotely but not locally—without pulling developers out of flow.

• His roughly six-month focus is seamless movement across mobile, GitHub.com, the web, and a local editor. The memorable framing is AI sprinkled into native workflows “salt-bae style,” with Agent HQ and GitHub co-evolving rather than operating as an isolated agent dashboard.

• Jared calls dev containers an important lightweight sandboxing concept and asks whether they should become a standard. The host points out that they can bundle VS Code, a filesystem, sandboxing, security controls, and GitHub Enterprise connectivity, while also noting competing runtime approaches inside Microsoft.

• Jared identifies repository setup as Cognition’s number-one pain point, and says Codex and presumably other agents face it too. He notes that a Next.js specialist can often assume npm install; the host explains that a general system cannot predict hidden dependencies such as ffmpeg. The discussion points to autodetection, preemption, and shared defaults, while the host recalls an unrealized idea for an open-source framework detector that could move the ecosystem together from npm to Bun.

4. Better agents require reliability engineering, not just smarter models

• Asked what follows the progression from Copilot to more autonomous coding agents, Jared rejects the hunt for a dramatic new category: “Making them good.” Moving from 90% to 95%, 98%, 99%, and eventually “nines of success” becomes progressively steeper, and “there is a big difference between 98% and 99% correct.”

• His warning is that many AI teams are “living in La La Land” about quality because they do not measure error-free sessions, provider failures, dropped requests, or latency. At Vercel, key metrics rolled up about every three hours—an almost “video game style” operating loop that made a good or bad day immediately visible.

• Infrastructure reliability is part of model quality because inference does not yet behave like a database with dependable uptime. Providers differ in both performance and availability and “go down all the time”; Jared connects the success of OpenRouter and gateway products to the need for switching and failover during multi-turn agent sessions.

• Computer use remains a modality under discussion: Jared says it became known as slow, bad, and inaccurate but is improving, especially with open vision models such as DeepSeek-OCR and olmOCR. The host calls it computationally intensive. Jared’s emphasis remains the hard final mile: “The devil’s in the details.”

• Jared also applies agents to personal workflows. He pointed Claude Code at his father’s accounting spreadsheets; it generated Python scripts and seemed better than ChatGPT, though it “got off the rails.” He uses agentic browsers, currently mostly Atlas while retaining Arc for its vertical tabs, and built Chrome Dump to summarize and close hundreds of tabs through Markdown. The host is surprised that data-analysis agents remain relatively underdeveloped, with organizations still in the BI era rather than querying precise analytics through a Slack bot.

5. Core GitHub workflows are being reopened for repair

• Jared cites an RZ tweet from May 22, 2025 declaring that “pretty much the entire GitHub homepage is useless”; it drew 1.3 million views and 19,000 likes. The new homepage responds with tasks, recent pull requests, and retained elements such as recent repositories. Jared praises the overhaul while stressing, “There’s more work to do.”

• Stacked diffs are the larger workflow fight. Jared’s analogy is a sequence of pull requests containing one commit each: developers can alter an earlier diff and restack the dependent work, with automatic restacking if the system supports it. The stack can then be collapsed and merged, and sophisticated setups can choose which layers deserve CI. For monorepos and very large codebases, that makes review and updates more fluid than conventional pull requests.

• The request has led GitHub’s community list for years. Jared found attempts reaching back to 2020 and a polished 2022 client-side implementation that introduced stacks outside the pull-request model; it was deemed too large and risky. The feature is now being worked into planning, but GitHub’s scale and Git implementation make it “not just a walk in the park.”

• That tension captures Jared’s operating posture: visible responsiveness without promising an easy ship. His DMs remain open because “all feedback is a gift” and “it’s all a signal”; collecting more of it should improve decisions about both Agent HQ and the less glamorous GitHub primitives underneath it.