Writing on AI agents, developer tooling, and software architecture — by Björn Roberg.
When xAI open-sourced their For You ranker on 2026-05-15, they redacted every numeric weight, every threshold, every Kafka topic, every config string — but they left every symbol name in source. For a serious competitor, the list of names is the more useful artifact. So either someone messed up, or that was always the plan.
Adding one optional frontmatter field to my blog produced a taxonomy I couldn't have written without doing the work. The categories come from Bereiter & Scardamalia (1987), and applying them to my own corpus was uncomfortable in useful ways.
MeMo claims near-SOTA prompting via Munger's mental models. Strip the framing: what's left is one trick — ask the LLM to name its reasoning strategy first.
Classical GP bundles five independent design choices behind one biology metaphor. Four of them are compute-era assumptions that don't survive contact with frontier LLMs. What's left is a different algorithm.
LLM generation costs roughly 1000× more than verification. The serious program-search systems already exploit this. The agent-scaffold mainstream doesn't — and nobody yet compounds search traces into a domain-specific mutator.
If Claude Code's Bash tool returns 'Exit code 1' with empty stdout and empty stderr for every command, including `true` and `echo ok`, your /tmp is almost certainly full. Here is the fix.
The AI-supply-chain cohort is already crowded: Invariant, Snyk Agent Scan, JFrog, Klaimee, a dozen more. They all ship the same artifact, a point-in-time report. That is not what underwriters can price against. The primitive insurance needs is a falsifiable attestation with a TTL that re-verifies on drift. Nobody ships that yet.
The biggest breakthroughs come from connecting unrelated fields — but the mechanism that produces them is broken. Hidalgo's firmbyte concept explains why.
Applying IoC and dependency injection patterns to agentic service boundaries — and why technical plumbing alone won't prevent control inversion.
Applying signal processing concepts to coding agent actions — from FFT and Markov models to a pragmatic health scoring algorithm for Claude Code hooks.
slog is a zero-dependency structured logger for Node.js, Deno, Bun, and Cloudflare Workers with pluggable transports and a Hono waitUntil middleware.
cairn is an append-only SQLite event store where immutability is enforced by triggers, not convention. Go, TypeScript, and Rust SDKs share one test spec.
The blog has moved to bjro.dev — a short domain that's easier to share and signals a long-term commitment to writing about software and technology.
Every Node.js production system depends on a single corporate-controlled npm registry with no real alternative — a structural risk the ecosystem ignores.
Debian turns 33 and keeps earning its title as the universal OS — community governance, 59,000 packages, and the foundation of over 1,000 Linux distributions.
Apply partial application from functional programming to AI agents — pre-configure agent parameters to create reusable, composable agent building blocks.
Linear types applied to AI agent safety — a type system that enforces single-use for destructive operations and statically rejects unsafe agent code.
MarkyMarkov uses Markov chains to learn patterns from your codebase and provide fast, explainable code guidance that complements LLM agents.
Philosophically informed AI agent design — modeling belief, knowledge, and epistemic norms — makes agent behavior more reliable and inspectable.
Combine Markov chains with LLMs to build structured, interpretable agents — Markov chains enforce predictable state; LLMs provide semantic flexibility.
How to pair agentic workflows with Temporal for durability, finite state machines for structure, and MCP servers to scale from pilot to production.
Context is the most underrated lever in AI agent design — this post makes the case for MINDMAP-style context objects as a first-class agent primitive.
Infra teams quietly feed logs, configs, and schemas to cloud LLMs when troubleshooting — extending the trust boundary into a black box they don’t control.
Opinionated notes on software development in the age of agentic tooling — where LLMs genuinely change how we work and where the hype outpaces the reality.
casq brings content-addressable storage to CI pipelines, eliminating redundant builds and test runs by identifying identical inputs with a git-style hash.
mise pins tool versions per project — Node, Python, Go, and more — in a single .mise.toml, replacing nvm, pyenv, and Makefiles with one unified tool.
whosthere is a Go-based LAN discovery tool with a modern terminal UI — fast, intuitive network exploration that shows who and what is on your local network.
casq is a Rust CLI and library for content-addressed file storage — git-style deduplication without the overhead of a full version control system.
A quick comparison of tmux and zellij as terminal multiplexers — zellij wins on discoverability with visible, mnemonic keybinds right out of the box.
Practical techniques for sandboxing AI agents and limiting blast radius — filesystem restrictions, network controls, and permission scoping in production.
Remap Caps Lock to Left Control for a more ergonomic keyboard layout — instructions for Debian/Linux via XKBOPTIONS and Windows via SharpKeys.