Claude Code cross-session messaging lets your sessions message each other. How ListAgents and SendMessage work, inbound controls, and the limits.

Eighteen agent definitions went from 8,798 lines to 664 and got better. What died: persona theater, THINK HARD, emoji. What survived, and why.

Anthropic cut 80% of Claude Code's system prompt. We ran the same surgery on a production framework: 8,798 agent lines to 664. Here is the diff.

Exit code 2 blocks an agent and feeds it a correction. Building a real TeammateIdle gate exposed three loop-safety rules the docs do not carry.

Claude got it wrong. Model problem or effort problem? The diagnostic that tells you which dial to turn, plus the effort field on every subagent.

An agent that finished and one silently blocked look identical. The two delivery models, why self-reports are not evidence, and what to check.

Every CLAUDE.md guide tells you what to add. After Claude 5, subtraction wins. The delete list, the one test per line, and what must never be cut.

Fable 5 rewards a briefed outcome and punishes step-by-step prompting. Six practices for Claude Code, plus where Opus 5 still wins.

When Claude Code multi-agent orchestration saves money and when it adds a 60% markup for nothing. The coordination cost math, measured.

Claude Opus 5 vs Fable 5: Opus 5 wins nearly every eval at half the price. Where Fable 5 still leads, the retention gap, and which to pick.

Claude Opus 5 benchmarks, pricing, and specs: $5/$25 unchanged from Opus 4.8, near-Fable 5 results, thinking on by default, and a new max effort.

A hidden Claude Code subagent watches another agent and reports drift. Observer agents: the enable flag, the digest, and the ObserverReport tool.