# Specialists Kady can delegate focused assignments to independent Pi agents in the same project sandbox. Ask for a named specialist, or let Kady select one: > Use the statistical-reviewer to check `results.ipynb` and report the checks performed. ## Available roles The scientific roster covers code/computation, literature/fact checking, study design and writing. Examples include `statistical-reviewer`, `citation-checker`, `peer-reviewer`, `data-validator`, `reproducibility-auditor`, `investigator` and `comparative-reviewer`. Specialists added in a later release appear once in existing projects; deleting one keeps it deleted. **Settings → Specialists** shows the installed roster and enabled state; the default personas live in [`subagents.ts`](../server/src/agent/subagents.ts). The delegation package also supplies general-purpose agents and external-CLI agents for Claude Code, Codex and Cursor. **External-CLI agents are disabled by default** because their usage bypasses Kady's runtime, ledger and spend cap. Native specialists can use files/shell, web, notebook, PDF annotations, Modal and [MCP tools](mcp-servers.md), subject to the parent's tool ceiling and their own allowlist. They do not receive the lead's interview form. ## Assignments and supervision A brief should name inputs, allowed edits, outputs, completion criteria and limits. Parallel workers share files, so give them distinct responsibilities. Handoffs report completed/partial/blocked status, supporting evidence, checks performed, artifacts and remaining questions. A completed review can find an invalid result; completion is not a scientific verdict. To run several specialists, or one in stages, Kady writes a short workflow script in its reply and launches it. The script appears in chat as a collapsed **Workflow script** block naming the specialists it starts; expand it to read the plan. Kady checks the specialists and models it names against the spend cap before anything runs. Kady can hand the chat back while specialists keep working; when they finish, it picks their results up in a new turn of the same chat, which an open tab shows automatically. A specialist uses `contact_supervisor` for blocking decisions. Kady resolves the request from existing instructions or asks you through the interview form. The child waits up to ten minutes; timeout/dismissal is not approval. Open **Automation → Specialist fleet**, select a chat and inspect live transcripts, models, tokens and activity. Controls let you steer, stop or resume work. ## Verification Reviewers check work in two modes. An ordinary review reports evidence-backed findings. A **verification gate** decides whether a result is accepted: Kady opens the brief with "Verification gate:", and the reviewer treats every step as unestablished and every citation as wrong until it has checked them itself, then ends with `Verdict: accept`, `repair` or `reject`, naming the first failing step and listing the steps it did confirm. Kady accepts a gated result only when its reviews pass and does not overrule an objection on plausibility. `comparative-reviewer` reads several independent attempts at the same question side by side. It catches what a per-attempt review misses: assumptions or inputs every attempt shares without justification, contradictions between attempts, and agreement that only reflects a shared flawed input. `/prove-verify ` runs rounds of this (after Cogentic, a multi-agent harness for proof discovery, [arXiv:2609.40324](https://arxiv.org/abs/2609.40324)): 1. Kady states the target precisely and creates `derived/prove-verify/-