Coding agents produce causal DAGs, not logs

I've been building tracing hooks for coding agents — Claude Code, Codex CLI, Copilot, and others. The goal was simple: record what these agents do, so I could debug them when things went wrong.

What I found surprised me. The standard observability abstraction — a flat, chronological log of events — is the wrong primitive for coding agents. These agents don't produce meaningful timelines. They produce causal DAGs.

This post explains why, shows the difference with real traces, and argues that runtime observability for AI coding agents needs a fundamentally different data model.

1. A debugging story

A Claude Code session ran 87 tool calls over 12 minutes. The agent was asked to fix a bug.…