Daily Digest — 2026-08-27

The speaker, Dr. Stephanie Jearue, a research scientist at Sourcegraph, begins by humorously acknowledging her transition from astronomy to AI research and, specifically, to the role of an "Agent Advocate." She sets the stage by tracing the evolution of developer advocacy (DevRel) from "software evangelism" in the 1980s to its more recent role in fostering bottom-up adoption and influencing product strategy through a feedback loop with developers. However, the core premise of her talk is that the very definition of a "developer" is changing, leading to "The Death of Developer Advocates" and the rise of the "Agent Advocate."

Dr. Jearue emphasizes that developers stopped working alone, and "developer" stopped naming one person. She points out two key shifts: Firstly, at a high level, engineers became orchestrators. With AI, engineers can delegate work, but not the accountability. AI fluency is now a hiring requirement for every role, and the bar for engineers is rising. Secondly, non-engineers became builders. Support, research, data science, marketing, and sales teams can now pick up APIs and CLIs that once required an engineer, often achieving quick starts without extensive training.

The central problem Dr. Jearue identifies is that an agent sits in two seats at once: 1. A user of your tools: An agent reads documentation, authenticates, calls APIs, runs CLIs, and recovers from errors. It interacts with tools much like a human developer, but as a machine, its interaction patterns and failure modes are different. 2. A recommender of your tools: Agents pick libraries, scaffold frameworks, and tell developers what to install. This means that agents are now influencing the adoption of tools, similar to how human developers recommend tools to their peers, but with potentially different motivations or biases.

To illustrate these points, Dr. Jearue shares concrete examples from her work building a benchmark called Code Scale Bench at Sourcegraph. This benchmark involves developing hundreds of tasks reflective of the software development lifecycle and unleashing agents with and without Sourcegraph's code navigation MCP tool. This allows for observing how agents interact with tools, encounter errors, and recover from them.

Measuring Agent as User (Seat One): She notes that agents kept making the same mistake because "it was our fault." Her analysis of 2,514 recorded agent sessions revealed that agents read error messages and fixed themselves, but often at the cost of "one wasted turn" thousands of times over. This highlights a crucial insight: while agents can learn and adapt, persistent errors point to flaws in the tool descriptions themselves. For example, an agent might try to use `readline` instead of `startline` for a tool, failing repeatedly. This underscores the need for precise and unambiguous tool documentation and design to minimize agent friction.

Measuring Agent as Recommender (Seat Two): Dr. Jearue explains that "being mentioned is not being recommended." In her Generative Engine Optimization (GEO) baseline experiment with Claude, 55.1% of responses mentioned Sourcegraph products across 1,610 responses (322 purchase-journey prompts, 5 repetitions each). However, only 38.8% of these were explicit recommendations (top-three placement or an explicit endorsement cue). More critically, 89% of failing prompts (those describing a specific pain point our product addresses) resulted in zero mentions, with the model recommending generic tools like `grep` or `ripgrep` instead. For example, when prompted with "We keep breaking downstream services when we change shared libraries because we can't see all the consumers" (a problem Sourcegraph's observability layer solves), Claude suggested making a wiki page, not our product. This reveals a significant gap in product messaging and discovery within the agent-driven recommendation landscape.

Guiding the Agents in Seat Two: To address this, Dr. Jearue offers several strategies for guiding agents toward product recommendations: Correct the drift at read time: Training data can quickly become stale in the fast-moving AI field. Product teams need mechanisms (like `LLMs.txt` files, as seen with Stripe) to constantly feed agents up-to-date information, preventing them from pitching outdated or less relevant products. Give the agent something to quote: Agents need structured, verifiable information (reference tables, comparisons, schemas, current examples) that they can confidently present to users. They want to be authoritative and accurate. Be present where agents shop: Products need to be discoverable in MCP (Multi-Cloud Platform) registries, package managers, agent plugin, and skill marketplaces. If an agent encounters too much friction or a lengthy sales process, it won't recommend the product. Cover the symptom-first prompt: Content campaigns should directly address user pain points and use cases, ensuring that product descriptions are clear enough for agents to pick up on these connections. The goal is for your product to be surfaced when a user (human or agent) is in their moment of need.

The Blurring Lines of DevRel: Engineering, Product, or Marketing? Yes. Dr. Jearue acknowledges the ongoing debate about where DevRel belongs, stating that with agents, these lines have only become fuzzier. However, this also presents an opportunity: adaptability is the job description. She breaks down the "flavors" of agent advocacy: The Engineering Flavor: Build the agent's path. This involves building MCP servers, instrumentation, and retrieval loops to allow agents to interact directly with the product. Evals become crucial for identifying and fixing agent friction. The Product Flavor: Own the agent's experience. This entails defining AI experience rubrics and integrating agent experience reports and friction points directly into the product roadmap. * The Marketing Flavor: Own the recommendation. This involves GEO (Generative Engine Optimization) baselining to understand agent mention and recommendation rates, shaping content and messaging to improve product discoverability.

What Carries Over? The core tenets of DevRel remain vital, but each gains a "second audience": 1. Enablement: Educate developers AND the agents working beside them. This requires Agent-Friendly APIs, docs, and tools. 2. Community: Mold developers who now write with their own agent in the loop. This raises new questions about privacy and data shared between humans and their agents. 3. Feedback loops: Systematize what agents reveal about your product, rooting it into product and engineering to drive iterative improvements. 4. Credibility: Earn trust from agents and from humans, on very different terms. This means avoiding "Claude's slop" (agent-generated content that is subpar or biased) and focusing on high-quality, structured information that agents can reliably parse.

Dr. Jearue concludes by advocating for building the "curb cut" for AI agents. By serving the agent, the human path gets clearer too, as improvements for one user group benefit all. She encourages engineers to point a coding agent at their own docs and develop an Agent Experience Report, and for GTM-focused individuals to run buyer prompts through Claude/ChatGPT to measure mentions and recommendations for a quick GEO baseline. The talk itself is "agent-legible," with a QR code leading to a text transcript and toy repositories for practical implementation.

DeepSeek Harness is presented as a novel open-source agent harness that enables AI agents to be self-extending and deeply customizable. This framework distinguishes itself by treating 'everything as a plugin': models, tools, sandboxes, session storage, UI, and even the agent loop itself can be modified or swapped out.

The core innovation lies in the ability of agents running within the harness to dynamically rewrite and extend their own functionalities. This means the AI can generate and integrate new code for plugins, modify its own user interface components, and even alter its operational logic in response to prompts, all without human developers manually writing the code for these extensions.

Key features and capabilities demonstrated include: Dynamic UI Generation: The video showcases the AI adding a swimming 'whale' SVG to the web UI and embedding a playable Snake game directly into the conversation panel. This highlights the ability of the harness to completely rewrite the user interface on the fly. Custom Agent Presets: Users can prompt the AI to create new agent 'modes' or 'personas'. An example is the creation of a 'Code Review Mode,' where the AI is instructed to analyze a codebase, identify issues, and rank them by severity, automatically generating the necessary configurations and logic for this new mode. This demonstrates how agents can be tailored for specific, complex tasks without manual coding of the agent's persona. Research Mode: An agent can be tasked with entering a 'research mode' to fact-check claims in a document against real research papers. This involves the AI dynamically searching, judging, and annotating claims based on external literature. Local AI Lab: The harness can create a local AI lab environment that monitors system metrics like tokens per second and GPU memory usage while the AI is thinking, providing real-time insights into its computational activity. * Video Production Storyboarding: For creative tasks, the AI can act as a video production agent, planning shots and generating storyboards based on a video description.

Architectural Foundations for Robustness and Reversibility: DeepSeek Harness addresses the challenge of system stability when agents modify their own code by integrating reversibility deep into its design. The system employs a 'coat check' analogy: every action performed by the agent (e.g., changing code) is tracked, and the system automatically generates and stores 'clean-up instructions' (akin to a ticket) that can later be used to revert the action. This ensures that new components can be safely added and removed, and any changes are reversible, preventing the system from becoming unstable or 'falling apart.' This is formally expressed through the `trackΓ = (f, g) ↦ (γ, φ) ↦ (f(γ), φ ◦ g)` mathematical representation, where `f` is the action and `g` is the cleanup instruction, and `φ ◦ g` ensures that cleanup actions compose correctly.

Efficiency and Cost-Effectiveness: Performance benchmarks presented in the video indicate that DeepSeek Harness achieves a comparable success rate to other agent frameworks (like Pi Agent, Claude Code, Open Code, and Hermes Agent) while being significantly more cost-effective. It boasts a cost per successful operation of $0.028, making it highly efficient for iterative development and experimentation.

Open-Source and Local Execution: DeepSeek Harness is designed to be open-source and can be run locally on a user's machine or on cloud platforms like Lambda. This commitment to open-source and local execution ensures privacy, control over resource usage, and eliminates concerns about token limits or external tracking. The framework's ability to run on various NVIDIA GPUs allows developers to train and fine-tune their own models or perform inference for tasks like text-to-image/video generation quickly.

OpenAI has officially unveiled its first custom AI chip, codenamed Jalapeño, designed specifically for AI inference. Surprisingly, this first-generation chip has achieved industry-leading speed and efficiency, outperforming established competitors like Nvidia's GPUs and Google's TPUs on critical metrics. The most important metric highlighted is performance per watt, which measures how much useful AI work a system can complete per unit of power consumed.

One of the most striking aspects of Jalapeño's development is the role of AI itself. OpenAI states, "We used AI to design the chip, and designed the chip so AI could program it." AI played a direct role in Jalapeño's development, accelerating the process from initial design to tapeout in just nine months. This was achieved by exploring implementations, shortening design, measurement, and verification loops, and continuously iterating on model workloads. AI also optimized the chip's arithmetic circuits, fitting more compute performance into the chip on schedule.

SemiAnalysis, an independent analytical firm, verified OpenAI's claims through their own testing using the InferenceX benchmark. Their report, titled "OpenAI Jalapeño: Better Than Nvidia Blackwell," confirms that Jalapeño indeed "smokes every other chip" in terms of token throughput per All-in utility MW. This means Jalapeño offers superior performance per watt across almost all scenarios, excelling in both low-latency and high-throughput scenarios without being specifically tuned for any particular point in the curve. An "apples to apples" comparison against Single Token Prediction results shows it "knocks every competitor out of the water."

Jalapeño is an ASIC (Application-Specific Integrated Circuit), meaning it's designed for a singular purpose: running large language models (LLMs) for inference as fast and efficiently as possible. Unlike general-purpose GPUs, ASICs are highly specialized. This specialization allows them to achieve extreme efficiency for their intended task. The chip's primary function is to handle the outputs of AI models, which involves complex mathematical operations, particularly matrix multiplication and attention mechanisms.

A key insight from the SemiAnalysis report concerns the software stack that interfaces with the hardware. Running an AI model effectively on a chip requires highly optimized kernels (low-level code that tells the chip exactly how to perform computations). Historically, writing efficient kernels is notoriously difficult and requires significant human expertise. Nvidia, for instance, has developed CUDA, a robust programming platform with decades of development, extensive libraries, tools, and a large ecosystem of engineers. This "CUDA moat" has long been considered a significant barrier to entry for competitors.

However, OpenAI has adopted a different approach. They developed their own kernel programming language called Gluon. What's truly remarkable is that OpenAI's AI models, specifically Codex, were used to write functional and efficient kernels for Jalapeño so quickly, circumventing the need for extensive human engineering intervention. This implies that AI is now capable of generating highly optimized, low-level code directly for hardware, a task previously reserved for highly specialized human engineers. By leveraging AI to write assembly-level code (the most foundational and manual level of programming), OpenAI is effectively automating the creation of highly efficient kernels, rapidly closing the software gap that has protected incumbents like Nvidia.

This capability, often referred to as recursive self-improvement or a "flywheel effect," means that better AI can design better chips and write better software for those chips, which in turn accelerates the development of even better AI. This vertical integration of AI designing both its hardware and software is a paradigm shift. Rumors suggest that OpenAI's next pretrain, "Bel" (successor to "Doug"), could have over 10 trillion parameters and potentially serve as the base for an AGI-threshold model. The flexibility of Jalapeño's architecture further supports future large-scale models, accommodating 10-20 trillion parameter models or 2-4 million token context windows. This signals OpenAI's ambition to control the entire stack, from foundational models to the underlying silicon and software, potentially disrupting the current landscape of AI hardware and software development where Nvidia has long held a dominant position. Sam Altman's concise tweet, "we made a chip and it is fast," encapsulates the groundbreaking nature of this achievement, highlighting the rapid, AI-driven progress in the field.

A novel AI coding model, Ox Alpha, has recently emerged on OpenRouter, drawing significant attention from the developer community. What makes this model particularly intriguing is its mysterious origin: OpenRouter states that it was built and is operated by an anonymous third-party provider, with no publicly known maker.

Despite the anonymity, developers have been actively experimenting with Ox Alpha, and the feedback has been overwhelmingly positive, with many describing it as "super impressive." This has sparked widespread speculation across platforms like X, Reddit, and Hacker News, as users attempt to "fingerprint" the model to identify its creators. Hypotheses range from major players like JLM or Google to Chinese AI companies or other US-based firms, but as of now, nothing has been confirmed.

Beyond the mystery, the engineering setup of Ox Alpha itself is noteworthy for senior software engineers. It is explicitly designed for coding and sustained agentic work. Key technical specifications include:

A 1 million token context window: This enormous context size allows the model to process extremely large codebases, extensive documentation, or prolonged conversations without losing context, which is critical for complex software development tasks. Multimodal input: Ox Alpha accepts not only text but also images and video. This capability suggests potential applications in areas like understanding UI mockups, analyzing video demonstrations of bugs, or interpreting diagrams and architectural blueprints directly. Tool-calling support: This feature enables the model to interact with external tools, APIs, and environments. For agentic engineering, this is a foundational capability, allowing the AI to execute code, run tests, query databases, or perform other actions to achieve its goals, moving beyond mere code generation to actual task completion. Currently free: The model's availability without cost on OpenRouter means that developers can run extensive agent trajectories and perform significant computational work using somebody else's model infrastructure without immediate financial overhead.

This "free experiment" aspect, coupled with the anonymous provider, presents a double-edged sword. While it democratizes access to a powerful coding AI, OpenRouter explicitly states that the provider retains your prompts and completions. Although they claim this data is not used for training, this policy raises significant concerns about data privacy, intellectual property, and security.

For agentic engineering, this situation serves as a critical reminder: model capability is only one component of the model selection process. Once autonomous agents can read entire code repositories, execute long-running tasks, and accumulate vast amounts of context, who operates the endpoint and what happens to the agent's trajectory (i.e., its sequence of thoughts, actions, and observations) may matter just as much, if not more, than the model's raw coding ability. The trust in the underlying service provider, their data retention policies, and their security practices become paramount considerations for any serious production deployment.