Delegation strategy

Delegation strategy

Pick the lowest rung on this ladder that fits the task. Every step up costs latency, tokens, and context-transfer overhead: a subagent starts with zero knowledge of the conversation.

The ladder only goes as high as your harness's tools. Before climbing, check what you actually have — subagent spawning? parallel execution? a workflow/orchestration tool? — and treat your highest supported rung as the ceiling. If a rung is missing, see "Degrading gracefully" below.

The ladder

1. Do it yourself (default) — you know where to look, the work is linear, or it's small.
2. One subagent — the search is wide but the answer is narrow, or a specialized agent type fits.
3. Parallel subagents — multiple genuinely independent strands of work.
4. Workflow orchestration — structured fan-out at scale, AND the user explicitly opted in.

When a subagent earns its cost

The core trade-off is context economy vs. directness. An agent burns its own context and returns only its final message.

• Delegate wide-search / narrow-answer work. "Which of these 40 files handle auth?" — let an Explore agent read the file dumps and hand back the conclusion. Your context stays clean for the real work.
• Never delegate single-fact lookups. If you already know the file, symbol, or value, a direct Grep/Read beats spawning an agent and writing it a prompt.
• Write the prompt as if to a stranger. The agent knows nothing you haven't told it: include file paths, constraints already discovered this session, and the shape of answer you want back (a list, a verdict, a path). Most delegation failures are under-specified prompts, not wrong rungs.
• Fan out independent work. Launch unrelated investigations as multiple agents in a single message so they run concurrently. Parallelism pays when the strands are slow or numerous, not merely independent — two quick lookups can stay sequential.
• Isolate parallel writers. Agents that edit files concurrently need disjoint file sets, or per-agent worktrees if your harness offers them. Logical independence isn't enough to stop them clobbering each other's changes.
• Match specialized agent types when your harness defines them (for example Explore, Plan, docs, security, or custom agent definitions). Their tools and prompts are scoped to the job; prefer them over general-purpose.
• Don't duplicate. Once you've delegated a search, don't also run it yourself. Wait for the result.
• Continue, don't respawn. If your harness can message an existing agent, follow up with it because it keeps its context instead of starting a fresh one.

When workflow orchestration is justified

This rung means a dedicated orchestration tool that runs many agents under deterministic, code-driven control flow. Two gates, both required:

1. Explicit user opt-in — a hard rule, not a judgment call. Only run workflow orchestration if the user asked for multi-agent orchestration in their own words, used an opt-in keyword your harness defines, invoked a skill that calls for one, or named a saved workflow. A task that would merely benefit from a workflow does not count. Without opt-in: describe what a workflow could do and its rough cost, and let the user choose.
2. The task's shape needs deterministic orchestration — control flow that should be code, not model judgment: fan-out over a known work-list (migrations, audits), independent finders + adversarial verification of every finding, loop-until-dry discovery, judge panels over competing designs. Scout inline first to discover the work-list, then orchestrate over it.

If the work is a single investigation or a linear edit, a plain subagent (or rung 1) is correct even when workflows are available.

Degrading gracefully

When your harness lacks a rung, translate the principle, not the tool:

• No workflow tool: emulate rung 4 at rung 3 — decompose into batches of parallel subagents and iterate, with you as the orchestrator. The opt-in gate survives the translation: spawning dozens of agents is a scale decision the user must make explicitly, no matter which tool does the fan-out.
• No subagents at all: rung 1 is the whole ladder. The principle becomes context hygiene — read narrowly, summarize findings as you go instead of retaining raw file dumps, and drop intermediate material once distilled. When a task is genuinely a fan-out job your tooling can't express (a 200-file migration, an exhaustive audit), say so and propose splitting it into sessions or steps the user drives, rather than grinding through it badly.

Failure modes to avoid

• Delegating trivial lookups → pure latency for nothing.
• Doing giant multi-file sweeps inline → context pollution that degrades the rest of the session.
• Escalating to a workflow without opt-in → token surprise; the scale is the user's decision, never inferred.