ways-of-working

Claude Code subagent orchestration manages multi-agent development by enforcing routing guidance, parallelism limits, and cost control guardrails. This ensures workflows stay predictable and prevents scope creep during complex repository tasks.

Adversarial review loops in agent orchestration are used when you need strict quality control during building and reviewing phases. They deploy critic agents to evaluate work, ensuring reliable execution and catching failure modes before task completion.

Routing tasks between parent-direct execution and parallel subagents depends on task complexity. The routing guidance evaluates whether to work inline, involve explorer agents, or fan out to parallel subagents based on the specific repository requirements.

Agent orchestration supports stall detection and failure-mode handling for repository tasks by applying execution discipline. It enforces retry rules and brief writing to keep planning, building, and reviewing phases focused and predictable.

Workflow routing can enforce brevity rules and cost limits across Claude agent teams. It applies selective subagent support and execution discipline to maintain quality control and prevent excessive token consumption during development.

Parallel fan-out for software engineering tasks is limited by parallelism constraints and stall handling rules. While it accelerates research and building, excessive fan-out can increase costs and complicate failure-mode management without proper routing guardrails.