wandb/senpai/delegate-subagents

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. Use agent_status for one non-blocking snapshot and cancel_agents when work is no longer useful; do not poll.