Opus-powered design and planning agent. Use for non-trivial architecture decisions, multi-file refactor planning, hard debugging where the failure mode isn't obvious, prompt engineering reviews on the custom agents (alphaseeker / energytransition / emergingtech), and any task where the main Sonnet agent would benefit…
Fast read-only file and codebase exploration. Use for "find where X is defined", "list files matching Y", "summarize what's in this directory", and similar grep/glob/read tasks. Cheaper than the built-in Explore agent (Sonnet-pinned) so the main agent can delegate liberally without burning Opus tokens.
Deep financial analysis subagent. Use for multi-step reasoning - DCF/intrinsic value, risk assessment, earnings analysis, adversarial bull/bear research. Operates on PII-filtered portfolio data and returns structured findings + recommendations. Never emits PII.
Read-only market-data and news fetch subagent. Use to gather prices, fundamentals, news, sentiment, and macro indicators for PII-free ticker symbols. Returns structured data ready to inject into analysis prompts. Does not reason or write files.
Read-only defect hunter for swarm workflows. Given a scope and a lens, finds real, evidenced defects and returns them as structured findings. Never edits files.
Implementer for swarm workflows. Takes a confirmed defect or an approved design and produces the minimal correct change, runs the relevant tests, and reports honestly.
Ship gate for swarm workflows. Reviews the current uncommitted changes against a task, runs verification, and returns a ship/no-ship verdict with concrete blockers.