# Bartłomiej Krupa

> Agentic Engineering · LLM Optimization · AI Cost Reduction. Bartłomiej Krupa engineers the harness around LLMs — agentic systems built for reliability and low cost. Deep dives, atomic notes, and a curated radar on LLM optimization, AI cost reduction, harness engineering, and agentic infrastructure.

## Profiles
- [GitHub](https://github.com/bartek-890): code and open-source work
- [LinkedIn](https://www.linkedin.com/in/bartek-krupa/): professional profile

## Deep Dives
- [Pick the right Claude tier for agent workflows](https://bartlomiejkrupa.dev/articles/claude-model-tier-comparison): Opus 4.8 orchestrates, Haiku 4.5 does the parallel grunt work, Sonnet 4.6 is the default in between. Choose by the job and the token bill, not the benchmark.
- [Context engineering beats a bigger context window](https://bartlomiejkrupa.dev/articles/context-engineering-beats-a-bigger-window): Claude Code runs up to a 1M-token window on flagship models, but a bigger window doesn't fix a cluttered one — every turn replays the whole history and quality drops before the cap.
- [Why agents ignore your CLAUDE.md](https://bartlomiejkrupa.dev/articles/why-agents-ignore-your-claude-md): A bloated CLAUDE.md gets filtered out as noise. Keep it to universal instructions and move everything else behind progressive disclosure.

## Garden
- [Keep CLAUDE.md to universal instructions](https://bartlomiejkrupa.dev/garden/claude-md-universal-only) (rule): Put only universally applicable instructions in CLAUDE.md; task-specific content raises the odds an agent treats the whole file as noise and ignores it.
- [Context engineering](https://bartlomiejkrupa.dev/garden/context-engineering) (concept): Context engineering is curating what enters the model's context window each turn, rather than relying on a bigger window to absorb the clutter.
- [Lost in the middle](https://bartlomiejkrupa.dev/garden/lost-in-the-middle) (concept): Language models use information at the start and end of a long context far more reliably than the middle — recall drops well before the token limit.
- [Subagent context isolation](https://bartlomiejkrupa.dev/garden/subagent-context-isolation) (concept): A subagent runs in its own context window and returns only a summary, keeping heavy intermediate tokens out of the parent conversation.

## Radar
- [SkillSpector](https://github.com/NVIDIA/SkillSpector): Scan before you install — 68 patterns, SARIF for CI, baseline suppression so re-scans surface only new findings.
- [learn-harness-engineering](https://github.com/walkinglabs/learn-harness-engineering): Harness engineering, not prompt engineering — 12 lectures and 6 projects that scaffold AGENTS.md, init.sh, and verification workflows.
- [codegraph](https://github.com/colbymchenry/codegraph): Returns line-numbered source plus call paths in one query — replaces a grep + read loop with a single round-trip.
- [claude-howto](https://github.com/luongnv89/claude-howto): Copy-paste templates and a progressive weekend path — useful in 15 minutes, not another feature reference.
- [Stagehand](https://github.com/browserbase/stagehand): act(), extract(), and agent() with action caching — AI when the DOM shifts, code when you already know the selector.

## Full Content
- [Complete raw text](https://bartlomiejkrupa.dev/llms-full.txt): every note and article concatenated for ingestion