Archives

After nine posts building a Claude Code-style agent by hand in Swift, I opened up JetBrains Koog framework and realized how much of what I'd hand-rolled was already someone else's problem. Here's what the framework gave me for free, and what it changed about how I think about agents.

The biggest gains in AI agent performance come from the system around the model, not the model itself.

How we restructured Claude Code's CLAUDE.md files in a monorepo to cut 73% of always-loaded context without losing a single fact.

Swift actors for background execution — the one type that truly needs an actor, with notification injection before each API call.

A file-based task DAG with dependency resolution — durable planning that survives context compaction.

A three-layer compression strategy — micro-compaction, auto-compaction, and manual compaction — for infinite conversations.

Two-layer knowledge injection — cheap awareness in the system prompt, full skill bodies loaded on demand via tool results.

Context isolation through recursive agent loops — spawning child agents with restricted tool sets for focused subtasks.

A Codable todo system with validation constraints and nag reminders — giving the agent a structured notepad it writes for itself.

Dictionary-based tool dispatch, path sandboxing, and the read/write/edit file tools that let the agent work with code.

The agent loop that drives everything — how a single stopReason check turns an API client into an autonomous agent.

Setting up a two-target SPM package, strict concurrency, and AsyncHTTPClient — the foundation for a CLI agent that never needs restructuring.