Archive

Writing

  1. Per-Call Authorization Cannot Bound an Agent Run

    An allowlist answers one question: is this single call permitted. The risk of an unattended agent accrues across the sequence, in spend, in repetition, in the fourth write to the same file. AWS shipping temporal conditions in a policy language is the first mainstream admission that a bound has to be stateful over a run, and the example they ship with it already has a concurrency bug in it.

    Systems
    Read
  2. The Pod Is the Wrong Unit for an Agent

    A Kubernetes scheduler knows two workload shapes and an agent is neither of them. The usual reading is that agents are bursty and idle, so pod-per-agent wastes money. The deeper problem is that an agent is three things with three different lifetimes, and a pod makes you give all three the same one. Anthropic, the Kubernetes Agent Sandbox project and kagent all unbundle those lifetimes, and they disagree about which piece keeps the pod.

    Systems
    Read
  3. Comprehension Debt Comes Due at Incident Time

    Agent-assisted teams are not shipping worse code. They are shipping code nobody on the team has had to be wrong about yet, and the process that used to build that understanding has been competed away against a peer shipping rate. A study of 26,000 students shows the shape of it, and Anthropic's own reliability engineers name the risk out loud.

    Systems
    Read
  4. Your Platform Has a New User Class, and It Does Not Use the UI

    The tools you buy have already rebuilt their primary interfaces for a reader that is not a person. Vercel shipped a language whose compiler output targets an agent, and Meta shipped a design system whose API, docs and CLI are designed together so a person and an assistant build the same way. Internal platforms are still assuming a human at a keyboard, and that assumption is what makes them legacy.

    Systems
    Read
  5. Your Model Provider Ships the Worst SLO in Your Stack

    The only Claude tier that ever named an uptime number targeted 99.5%, and it is no longer available for purchase. Everything else runs on documented best-effort availability. That is a weaker commitment than the cheapest tier of commodity compute, and almost nobody architects around it.

    Systems
    Read
  6. Agent Skills Are Speed-Running Package Management

    In four months skills went from loose markdown to registries, installers and a content-hash lockfile. My own machine has 67 entries in ~/.claude/skills and 61 of them are broken symlinks pointing at a source that no longer exists. The ecosystem is rediscovering npm's primitives in npm's original order, which puts provenance last.

    Systems
    Read
  7. Running Agents Is a Platform Problem

    Microsoft's Agent Harness reached GA the same week InfoQ published Deutsche Telekom's LMOS talk, and the two argue the same point from opposite ends of the market: the hard part of running agents is not the agent. In ShipMore, the product I sell on an agent-first story, the entire agent-facing surface is 703 lines against 22,769 lines of service layer behind it.

    Systems
    Read
  8. Exit 0 Is Not Success for an Agent Job

    OpenAI's agent hacked Hugging Face, ran for days, and went unnoticed for a week. For anyone running scheduled agents the lesson is duller than the security story: a run where half the sources failed exits 0, writes a file the same size, and looks identical to a good one. Two cheap fixes from a cron job I actually run: make the agent report its own coverage in the artifact a human already reads, then let deterministic code decide whether it ships.

    Systems
    Read
  9. Capacity Planning for Agents: Measure Yield, Not Tokens

    Goldman says 24x token growth by 2030. Others say 240x once you count parallel agents. Both are demand forecasts you can't spend against. Take a curation pipeline that reports keeping 12 items from 37 swept: follow the funnel to the end and 641 candidates become 202 published items, 5 things worth acting on, and 3 anyone actually did. The metric that matters is cost per accepted change, and it sits on the wrong side of every agent's telemetry boundary.

    Systems
    Read
  10. Agent Memory Wants a Wiki, Not a Vector Store

    LangChain is reframing agent memory as a hyperlinked wiki, and a logging study found Claude Code ships ~33k tokens of overhead before your prompt even arrives. I've been running memory as a citation-linked markdown wiki for months, no embeddings anywhere in the pipeline. Here's what the setup looks like and where it beats top-k retrieval.

    Systems
    Read
  11. Agent Sprawl Is an Identity Problem, Not a Dashboard Problem

    Google's second State of AI Infrastructure report says 83% of orgs need to upgrade their infrastructure for agents. The number that actually matters is that nobody can see what their agents can touch, and a governance dashboard sitting on top of a wide-open credential is decoration. The fix is scoping the credential.

    Systems
    Read
  12. The Sandbox Problem for AI Agents: Lambda MicroVMs and Where Untrusted Code Should Run

    An agent writes code you never reviewed and then runs it, so the question that decides whether it is safe to ship is the blast radius. Lambda MicroVMs give each session a Firecracker VM that resumes from a snapshot in under a second, including a pattern where Anthropic runs the agent loop and you run the tool calls.

    Systems
    Read
  13. Choosing an AI Agent Pattern: A Decision Guide for Enterprise Teams

    Most teams reach for multi-agent orchestration too early. Here is how to pick the simplest pattern that does the job, from a single direct model call up to a manager and its workers.

    Systems
    Read
  14. Graduated Trust for Production AI Agents: Parameterize Permissions by Phase, Not Environment

    Stop promoting AI agents across dev, staging and prod. Bind their permissions to the evidence they have earned, and leave the hostname out of it.

    Systems
    Read
  15. Maintainability in the Face of API Complexity

    The what, why, and how of API gateways.

    Systems
    Read
  16. Predicting Heart Disease from the UCI Dataset

    An end-to-end ML walk-through: wrangling four UCI cardiology datasets, exploring what actually moves the diagnosis, and pitting logistic regression against an SVM that lands ~85% cross-validated accuracy.

    Research
    Read
  17. Exploring Modern iOS Architectures

    MVC, MVVM, MVP and VIPER: the same TMDb movie app built four ways, compared on orthogonality, ease of use, and testability.

    Design
    Read