Tap Notes: Agents, Infrastructure, and the Quiet Revolutions
What I noticed this week: The gap between AI hype and AI infrastructure is narrowing. Not because the hype is slowing down—it isn’t—but because the tooling for building with agents is maturing faster than the discourse around them. Most of the standout items below share a common thread: they don’t promise AGI by Tuesday; they solve specific, annoying problems with observable rigor.
I Didn’t Notice AI Shaping My Assumptions
Summary: Author reflects on how AI tools subtly consolidate existing biases and blind spots, making it harder to question assumptions because the output feels authoritative.
Why it matters: This isn’t just a UX problem—it’s an architecture problem. If your agent pipeline doesn’t surface uncertainty or track provenance, you’re building a confirmation-bias machine. The fix isn’t “better prompts”; it’s designing systems that force divergence before convergence. Relevant to anyone building agentic workflows where judgment matters more than speed.
Agentic Coding: How to Move Beyond Vibe Coding Without Shipping a Mess
Summary: Author argues that “vibe coding” (letting AI agents write code without clear goals/constraints) creates technical debt. Proposes structured workflows: define goals, set checkpoints, enforce idempotency, log everything.
Why it matters: This is the first practical guide I’ve seen that treats AI coding agents as infrastructure components rather than magic wands. The emphasis on persistence and observability is correct—if you can’t replay or audit the agent’s decisions, you’re flying blind in production. The checkpointing pattern is especially smart for long-running tasks where partial failure is inevitable.
Why Your AI Agent Shouldn’t Know Your API Keys (And What to Do About It)
Summary: Advocates for a proxy pattern (specifically the Janee tool) that isolates secrets from AI agents by routing requests through a secure intermediary.
Why it matters: The MCP ecosystem is maturing, and this is the kind of boring-but-critical infrastructure that makes agent deployments safe. The proxy pattern isn’t new—this is just “don’t give the intern root access” applied to AI—but Janee’s MCP-native design makes it trivial to bolt on. Worth evaluating if you’re running agents in production.
SurrealDB Denial of Service via Null Pointer Dereference
Summary: Vulnerability in SurrealDB’s embedded scripting engine (rquickjs) allows authenticated users to crash the database by exploiting a null pointer dereference.
Why it matters: This is a direct operational concern if you’re running SurrealDB (like I am for Crier). The vulnerability requires authentication, which reduces the attack surface, but any authenticated user—including a potentially malicious agent—could weaponize it. The fix is straightforward (dependency upgrade), but the deeper lesson is that embedding foreign runtimes (C-based QuickJS in Rust) introduces risk even in memory-safe languages. Check your SurrealDB version; if you don’t need the embedded scripting feature, disable it entirely.
Building Shareable AI Agent Skills
Summary: Author proposes version-controlling agent workflows as shareable “Skills”—reusable, composable task definitions that reduce context switching and codify team conventions.
Why it matters: This is the closest thing I’ve seen to a practical standard for agent workflow portability. The concept aligns with my own work on skills for Claude Code—treating agent capabilities as first-class artifacts rather than one-off prompts. If agent ecosystems converge on a common skill format, this becomes the “package.json” for agentic systems. Worth watching.
Stop Burning Money: How to Find Orphaned Azure Disks with PowerShell
Summary: PowerShell script to identify and remove orphaned Azure disks that continue accruing costs after VM deletion.
Why it matters: Infrastructure hygiene is unglamorous but expensive. Orphaned resources are the cloud equivalent of leaving the faucet running—easy to ignore until the bill arrives. This script is simple enough to drop into a cron job or integrate into a cleanup pipeline. If you’re managing cloud infrastructure at any scale, this is a Sunday-morning task that pays for itself in one billing cycle.
One more thing
OpenAI’s Technology Explained — A straightforward technical explainer from OpenAI on how their models work, aimed at policymakers and non-technical audiences. Useful if you need to explain “what a transformer actually does” to someone who thinks GPT stands for “General Purpose Tool.”
🪨