Agentic Coding

Opus 4.8's parallel orchestration is still the biggest shift this week.

The dynamic workflows feature lets a single session spawn hundreds of subagents via JavaScript orchestration scripts. You trigger it by including "workflow" in your prompt or setting /effort ultracode. The built-in /deep-research workflow demonstrates the fan-out-and-verify pattern you'll reuse across documentation, security audits, and research tasks. Early testers report fewer tool calls needed and better self-correction before commits.

Try pairing it with yesterday's background agent mode for real throughput gains.

The 80% cliff keeps catching people mid-session.

When your token usage crosses that threshold, performance degrades faster than you'd expect. The fix: monitor with /context, use /compact to summarize without losing session memory, and schedule proactive restarts every 30 minutes on long tasks. The RALF loop pattern from earlier this week automates handoffs between sessions so large tasks complete without overloading any single agent's context window.

Build the exit-and-restart habit now—it compounds.