Writing
I got curious about how AI coding agents actually work on disk, not the chat UI, the real internals. So I took apart how Claude, Codex, and Antigravity store context on my own machine and wrote up the rabbit hole. A short series.
How Claude stores context
Sessions are JSONL conversation trees, messages link by parent UUID, tool outputs sit beside the conversation, and edits become versioned artifacts. It reads less like a chatbot and more like Git, event sourcing, and an IDE runtime combined.
18 reactions on LinkedIn
Reverse engineering AI agent internals
Claude and Codex are elegant: JSONL ledgers, SQLite indexes, worktrees, shell snapshots. Antigravity is something else, closer to an autonomous browser OS. The pattern: agents are quietly becoming IDEs, browsers, memory systems, and operating environments.
26 reactions on LinkedIn
Antigravity stores environments, not conversations
DOM snapshots, browser recordings, WebP playback artifacts, task brains, walkthrough memories. It does not just store what was said. It stores the whole environment it was said in.
14 reactions on LinkedIn
AI writes code. It is still figuring out design.
A few months of encoding my own preferences on spacing, color, and type into something an agent can follow. Most design systems add iteration loops instead of removing them. This one tries to remove them.
19 reactions on LinkedIn