Skip to content

wandb/senpai / plugins/senpai/skills/delegate-subagents

delegate-subagents

Verified

Choose, configure, launch, and collect bounded subagents for research and engineering decisions. Root agents and delegation-capable subagents should read this before delegating: every task requires an explicit model tier, and high-leverage work such as research ideation, round planning, plateau pivots, large research reviews, hard optimization, disputed evidence, and expensive experiment portfolios requires frontier judgment.

FileKindSize
SKILL.mdskill md5.0 KB
agents/openai.yamlconfig206 B
---
# SPDX-FileCopyrightText: 2026 CoreWeave, Inc.
# SPDX-License-Identifier: Apache-2.0
# SPDX-PackageName: senpai

name: delegate-subagents
description: >
  Choose, configure, launch, and collect bounded subagents for research and
  engineering decisions. Root agents and delegation-capable subagents should
  read this before delegating: every task requires an explicit model tier, and
  high-leverage work such as research ideation, round planning, plateau pivots,
  large research reviews, hard optimization, disputed evidence, and expensive
  experiment portfolios requires frontier judgment.
---

# Delegate subagents

Batch independent tasks in one `spawn_agents` call with a stable `batch_key`. Give each child a self-contained assignment and ask for a compact, evidence-linked result.

Every task must explicitly set `model` to `fast`, `smart`, or `frontier`; there is no implicit model tier. Choose the least expensive tier that provides the judgment the task requires:

- Use `model="fast"` for mechanical, easily verified work: locating files, extracting facts, counting records, formatting results, running bounded commands, and reporting test or build failures.
- Use `model="smart"` for ordinary implementation and review, literature retrieval, bounded synthesis, standard failure diagnosis, and debugging code that is not already heavily optimized.
- Use `model="frontier"` for quality-first research judgment. This includes fresh research ideation, planning a new research round, changing direction after a plateau, reviewing a large research or experiment history, difficult debugging or optimization of code that is already highly optimized, reconciling conflicting evidence or disagreement between local and external evaluation, and selecting a portfolio that will consume substantial GPU time or external-evaluation budget.

Keep frontier tasks focused. When using more than one for the same decision, give them distinct questions or perspectives. Do not spend frontier capacity on routine monitoring, simple retrieval, formatting, or other work whose answer is cheap to verify. Treat every child result as advice: inspect its evidence before acting on it.

## Choose an agent and context

Choose:

- `explore` for local code, data, artifacts, or history;
- `search_general_web` for current public sources;
- `search_research_publications` for scholarly literature and primary papers;
- `bash-runner` with `model=fast` for tests, builds, and bounded commands; and
- `general-purpose` for mixed analysis, planning, review, or implementation.

Agent specialization and model tier are independent. For first-principles synthesis, critique, diagnosis, or planning, use `agent="general-purpose"`. Use `search_research_publications` when the task is to find and compare primary papers, and `search_general_web` for current public sources. The search agent and its search skills own source selection and search mechanics.

Normally set `include_context=false` and provide a self-contained task with exact evidence paths. This gives the child a fresh perspective while preserving access to the merged system prompt and searchable parent history. Set `include_context=true` only when the complete model-visible conversation is necessary and cannot be summarized reliably. For research judgment, ask for research, critique, diagnosis, ideas, or a plan rather than edits.

## Keep bulk output in the child

Delegate before running a command or broad read that may return substantial
output. A child can summarize output it produced; it cannot remove output that
already entered the parent context.

- Use `agent="bash-runner", model="fast"` for verbose tests, builds, bounded
  logs, Git inspection, and deterministic command output.
- Use `agent="explore", model="fast"` for broad file, history, or PR-artifact
  searches and extraction into cited facts.
- Use `agent="general-purpose", model="fast"` for exact mechanical edits with
  explicit file targets and focused acceptance checks.
- Use `model="smart"` for ordinary implementation, review, synthesis, and
  failure diagnosis.
- Use `model="frontier"` for causal scientific interpretation, disputed
  evidence, direction changes, and expensive experiment portfolios.

For a potentially large GitHub review, have the root call
`get_prs(max_inline_prs=0)`, pass the returned artifact path to an Explore
child, inspect the child's cited decisive evidence, and perform any typed
GitHub transition in the root.

## Collect results

`spawn_agents` returns task IDs immediately. Continue useful work, then use bounded `await_agents` calls with `all`, `first`, `quorum`, or `change` and a timeout of at most 300 seconds; a larger timeout is rejected, so wait again instead. Use `agent_status` for one non-blocking snapshot and `cancel_agents` when work is no longer useful; do not poll.

A child's result is the text it wrote with its final `finish` call plus the finish message. A result above about 15,000 tokens is saved to a file under the role state's `delegation/results/` directory, and the result carries a summary and that path; read the file when the summary is not enough.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall