Week 19, 2026 · May 4–10: Digests, Hermes Agent and the Self-Improving Loop

Week 19: Digests, Feeds and Watching My Own Build

What I learned

Building a Personal AI Observability Digest

  • I asked my AI assistant Lili to set up a weekly AI observability digest that searches the latest news every Sunday evening and sends me a formatted report with personalized action items. The first edition covered the Gartner XAI prediction that XAI will drive 50% of GenAI observability investment by 2028, Datadog’s State of AI Engineering report and Microsoft Agent 365 going GA
  • The Gartner prediction that caught my attention was not the 50% number but the metric shift underneath it. The industry is moving from measuring speed and cost toward factual accuracy, hallucination rates and logical correctness. Most teams I see are still optimizing for p95 latency. The teams that switch to quality metrics first will have a meaningful advantage
  • Datadog’s telemetry across thousands of customers showed 8.4 million rate limit errors in March alone. 70% of organizations now run 3 or more models in production. Fleet complexity is growing faster than operational maturity and I suspect many teams are adding models faster than they are retiring old ones
  • Microsoft Agent 365 extends existing identity, security and compliance infrastructure to AI agents. This confirms what I have been telling my team: agent governance will be the differentiator, not agent capability. Enterprise buyers will ask “how do you govern this” before they ask “what does it do”
  • I automated the entire workflow via Lili with a cron job. Every Sunday at 8 PM IST, it searches the latest AI observability news, creates a formatted HTML and Markdown digest and delivers it to me. This is now running unattended

RSS Reader With 44 Feeds

Spotify API From Chat

  • I wired my AI assistant Lili to the Spotify Web API via OAuth. I can now say “play something”, “skip this track” or “search for that song” from chat. Nothing revolutionary in isolation but the meta-lesson is that every integration you automate is mental energy you get back
  • Lili handled the OAuth flow, token refresh and the JavaScript module that wraps the Spotify API calls. My job was to authorize the scopes from the browser. The OAuth scope management was the hardest part — I had to reauthorize three times because I kept missing scopes. Once it was set up, the refresh token flow handled itself
  • This is a good example of how I work with Lili: I describe what I want done, she builds it and I review and approve. The speed of iteration is the real advantage. A Spotify integration that would have taken me an evening took 20 minutes with her doing the wiring

Chrome Silently Installed a 4GB AI Model

  • This was the top story on Hacker News this week with 719 points and 530 comments. Google Chrome is auto-downloading Gemini Nano (approximately 4GB) without explicit user consent on the device
  • The backlash is not about the model itself. It is about the lack of transparency and the disk space impact on devices with limited storage. This is a deployment governance problem — as Malwarebytes covered — and I expect regulatory attention if Chrome continues this pattern without opt-in

Bun Is Being Ported From Zig to Rust

  • The Oven-Sh team is rewriting Bun from Zig to Rust. 673 points and 483 comments on Hacker News. The Register has a good summary. The discussion centers on the maintainability cost of Zig vs Rust for a project of Bun’s complexity
  • I believe the language choice matters less than the team’s honesty about the migration. It is rare to see a popular runtime admit that the initial language choice created long-term friction and then commit to a migration mid-stream, as Jarred Sumner explained on HN

Anthropic Published on Finance Agents

  • Anthropic published a paper on agents for financial services and insurance. The patterns for regulated industry agent deployment are directly relevant to the AgentForce and ServiceNow Assist work I do
  • Key takeaway from the Anthropic paper: Anthropic is positioning their model as a platform for regulated workflows rather than just a chat interface. The insurance claims handling example shows structured multi-step agent orchestration with human-in-the-loop at decision boundaries

OTel GenAI Conventions Are Moving to Their Own Repository

  • The OpenTelemetry LLM Semantic Convention WG announced this week that GenAI conventions are getting their own repository built on SemConv v2 and the Federated SemConv OTEP. A Python genai repo prototype is also being set up as a clean slate for native instrumentation
  • I started monitoring the Tooling WG and LLM WG meeting notes to track where the standards are heading. The May 5 LLM WG discussed agent planning spans as a new operation type and a GenAI SIG roadmap for the next quarter
  • The agent span conventions now define create_agent, invoke_agent, invoke_workflow and execute_tool operations — this is the schema my ragratch project needs to align with. I documented the full mapping of ragratch features to OTel conventions in my knowledge vault

What I built

  • Launched openodia.com — a portal for open-source Odia language tools, datasets and AI resources. The site covers three pillars: Education, Tooling and Community.
  • Recorded a video walking through how I built openodia.com and published it on the OpenOdia YouTube channel. This was my first video in over a year. The process was simpler than I remembered — AI handled most of the scaffolding
  • Lili shipped PR #5 on widescope — added a live demo link, OG meta tags for social sharing, a Quick Demo section and a Contributing guide to the README. She also fixed 5 clippy collapsible_match warnings that were breaking CI on the main branch. I reviewed and merged. This is how most of my open source work works now: I describe the change, Lili drafts the code and I verify and ship
  • Built a structured knowledge vault (AgentWiki, private repo) from 849 existing markdown files spanning the full range of my work and research. Lili and I crystallized 10+ reference pages in two days — approaches to harness engineering, OTel GenAI semantic conventions for agent spans, and an analysis of the Factory coding agents architecture. The vault has its own LLM-agent-readable AGENTS.md with content conventions (new-up-old-down order, Mermaid diagrams, auto-triage). This replaced scattered notes with a canonical knowledge base I can query directly
  • Gave Lili her own email address with Gmail API and Google Calendar API access via OAuth. The setup took about 10 minutes of my time — create a Google Cloud Project, enable APIs, authorize scopes via one URL click. She had full read-send-archive access within 15 minutes of me sharing the credentials
  • Set up an email watcher that checks my consolidated inbox every 5 minutes via cron. Forwarded emails from multiple mailboxes land in Lili’s inbox. She auto-archives newsletters, flags human emails to me on Telegram and stays silent at night. The 5-minute check cadence is aggressive enough to catch urgent emails within the same meeting slot without burning API quota
  • Created my first AI Observability Weekly Digest as a reusable format. The template includes 7 stories with deep-dive context, source links and personalized action items. Lili wrote the first draft and I refined the focus areas. I can now republish this every week via a cron schedule
  • Gave Lili permission to build a self-improvement system — a skills registry at /data/workspace/skills-registry/ where she stores reusable patterns from complex tasks, a session boot protocol that reads past learnings at startup and a weekly consolidation cron that reviews daily notes and updates long-term memory. This gets her closer to the autonomous learning loop that Hermes Agent has natively

What I failed

  • The Hacker News analysis I did for myself this week made me realize I consume far more content than I produce. I spent an hour browsing HN and summarizing it but I did not ship any of my own ideas during that time
  • I did not make progress on WideScope’s trace comparison feature. The code changes Lili delivered were documentation and CI fixes, not feature work. The comparison view is the feature that would make WideScope genuinely different from other trace viewers and I keep deferring it

Hermes Agent: The Self-Improving Agent

  • Nous Research released Hermes Agent, an open-source (MIT) autonomous personal agent with a built-in learning loop — it creates skills from experience, improves them during use, nudges itself to persist knowledge and builds a deepening model of the user across sessions
  • The architecture is interesting: multi-platform gateway (20+ messaging services from one process), model-agnostic (OpenAI, OpenRouter, your own endpoint), real sandboxing (Docker, SSH, serverless backends) and research-ready (batch trajectory generation, RL training environments)
  • I created a detailed wiki page in my private knowledge vault analyzing Hermes Agent. The Hermes Agent docs and GitHub repo cover the full architecture. My take: the self-improving skill loop is the feature that genuinely sets it apart, and the RagWatch instrumentation opportunity on Hermes’ sub-agent execution is worth exploring as a blog post
  • Lost a battle with Spotify OAuth scope configuration. Took three re-authorization rounds to get Spotify access tokens with the right permissions. I should have read the scope documentation first instead of guessing

Random Thoughts & Quotes

  • I believe AI observability is moving from a nice-to-have to a compliance requirement faster than most teams realize. The teams that treat eval sampling, XAI tracing and agent governance as product features now will not be scrambling in 2027
  • “When everyone has AI and the company still learns nothing” — This HN post (171 points) made me think about whether my own tooling is generating insight or just noise. A daily RSS digest is useful. A weekly AI observability roundup is useful. But both are consumption patterns. The real output is what I build from what I learned
  • A pattern I am noticing this month: the tools that impress me most are not the ones with the most features. They are the ones that get out of my way. blogwatcher, widescope and the Spotify integration all share one property: they do exactly what I need and nothing else. I want to keep this minimalism in everything I build
  • Giving an AI assistant its own email address turned out to be the highest-leverage integration I did this week. Email is universal. Every service supports it. No SDKs, no API docs to read. Once Lili had a Gmail account with API access, she could monitor my inbox, send on my behalf and integrate with anything that sends email notifications. This is the protocol that beats every integration pattern for reach