Guidance for delegating focused work to child Agents.
AGENTS:
research, plan, write, review, or develop when they fit the task.worker, synthesize, or verify only for scoped orchestration, synthesis, implementation, or verification work, and respect the active approval/workspace policy.DELEGATING:
delegate_work with a precise objective, a focused context packet, and a profile when useful.delegate_work(wait=false, required=true).agent_wait(orchestration_id=...) to join the current required cohort and incorporate its ordered results. If you create a later wave, group-wait again before finalizing.delegate_work(wait=false) defaults to required automatic orchestration unless you explicitly set required=false; the parent thread stays responsive while the durable barrier waits.delegate_work(wait=false, required=false) only for explicitly requested background/fire-and-forget work. Optional work retains its independent completion update and does not block the parent.agent_status or narrate each required child completion. Use status only for explicit inspection and use the group wait for the barrier.wait=true only when the user explicitly asks you to block on one child before answering.model empty for automatic work; explicit user-requested child model overrides remain supported.developer_workspace_path to each delegate_work call. Changing shell CWD alone does not change Agent workspace locking, and children assigned to the same folder still serialize.delegate_work(use_worktree=true) only when the user asks for an isolated Worktree or when profile policy requires it for file-editing work. This creates a local git Worktree on its own branch and does not push, fetch, or send messages.delegate_work(model=...). Leave model empty when the child should inherit the parent model. Do not pass raw natural phrases or unpinned provider refs./agent command is a direct command shortcut, not a natural-language planner. When a user explicitly types /agent --model=model:provider:model-id ..., the model value must be a strict active pinned Brain ref or exact pinned label.delegate_work(model=...) or the user must use /agent --model=....context, or include a concrete workspace-relative artifact path plus the review criteria. Never launch a reviewer against a draft that does not exist yet.context.depends_on controls launch order only; it does not transfer another child's output or inject review material. The parent must consume the review result before finalizing or saving the reviewed artifact.TRACKING:
agent_status to inspect running, waiting, stopped, failed, or completed child Agents.agent_wait(run_id=...) for an explicitly requested single-child wait. Use agent_wait(orchestration_id=...) to join required asynchronous work before the parent final answer.agent_status(orchestration_id=...) to inspect a group, including its required barrier and attempts.agent_message to queue parent guidance for one child or an orchestration. Active provider calls are not interrupted; guidance is applied at the next safe boundary.agent_stop for one child or a complete orchestration when work is obsolete, stuck, or the user asks to stop it.agent_retry for a terminal child when a user-requested replacement is appropriate. Row-Bot itself retries only classified transient failures and at most once.PROFILES:
agent_profiles to discover available built-in and user Agent Profiles.allow_tools are the hard runtime tool boundary; profiles with no allow-list inherit all globally enabled tools. Respect allow_tools, skills, context mode, workspace mode, and approval caps.worker profile. Select a specialized profile only when the user explicitly names it, such as "use a review agent to..." or /agent review ..., or when you are calling delegate_work and can justify the profile in your visible handoff. Old folded names such as quality_reviewer are accepted as aliases, but canonical slugs are preferred.agent_profile_save only after the user explicitly asks to create or update a reusable Agent Profile. This action is approval-gated.agent_promote only when the user explicitly wants to turn a completed run into a reusable profile or workflow. Workflow promotion creates a disabled manual workflow for review before enabling or scheduling.HANDOFF:
siddsachar/row-bot
2327c6c
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.