Delegation

Delegation — Agent Orchestration & Parallelization

What It Does

Parallelizes work across six patterns: built-in agents, worktree-isolated agents, background agents, custom agents via inline briefs, agent teams via TeamCreate, and parallel task dispatch. It also splits delegation into two weights — lightweight one-shot workers (capped turns) versus full agents that iterate with tools. The Algorithm auto-invokes it once work hits three or more independent workstreams.

The Problem

Doing independent work serially wastes time, but throwing every task at a heavyweight agent wastes more — spawning a full agent for a one-shot classification burns 10-30 seconds of startup for nothing. Worse, people conflate two different systems: fire-and-forget custom agents with no shared state, and persistent agent teams that message each other and share a task list. Pick the wrong one and you either over-coordinate simple work or under-coordinate complex work. This skill routes each job to the right pattern and the right weight.

How It Works

Auto-invoked by the Algorithm when work can be parallelized or requires agent specialization.

🚨 CRITICAL ROUTING — Two COMPLETELY Different Systems

These are NOT the same thing:

• Custom agents = one-shot parallel workers with unique identities, launched via Task(), no shared state
• Agent teams = persistent coordinated teams with shared task lists, messaging, and multi-turn collaboration via TeamCreate

When the Algorithm Should Use This Skill

• 3+ independent workstreams exist at Extended+ effort level
• Multiple identical non-serial tasks need parallel execution
• Specialized expertise needed (architecture design, implementation, ISC optimization)
• Large codebase changes spanning 5+ files benefit from parallel workers
• Research + execution can proceed simultaneously
• "Create an agent team" — use TeamCreate for persistent coordinated teams
• Unattended autonomous work where auditability matters more than speed — the Observer-team pattern (read-only agents watching the tool-activity audit log, voting continue/halt/escalate) is a fit here, but it is not currently implemented; it was retired with the Agents skill and would need to be rebuilt as its own component before use.

Delegation Patterns

1. Built-In Agents

⚠️ Built-in agents are for internal workflow routing ONLY. When the user asks for custom, specialized, or uniquely-voiced agents, write an inline brief per agent (section 4 below) and launch general-purpose.

Use Task(subagent_type="AgentType") with these specialized agents:

Always include: Full context, effort budget, expected output format.

2. Worktree-Isolated Agents

Run agents in their own git worktree with isolation: "worktree" for file-safe parallelism:

Task(subagent_type="general-purpose", isolation: "worktree", prompt="Senior engineer, TDD. ...")

• Each agent gets its own working tree — no file conflicts with other agents
• Worktree auto-created on spawn, auto-cleaned when agent finishes (unless changes made)
• Use when multiple agents edit the same files or for competing approaches
• Can combine with run_in_background: true for non-blocking isolated work
• Built-in agents with isolation: worktree in frontmatter auto-isolate on every spawn

3. Background Agents

Run agents with run_in_background: true for non-blocking parallel work:

Task(subagent_type="general-purpose", run_in_background: true, prompt="Senior engineer, TDD. ...")

• Use when results aren't needed immediately
• Check output with Read tool on the output_file path
• Ideal for: research, long builds, parallel investigations

3b. Foreground Agents (the default — contrast to pattern 3, not a seventh pattern)

Standard Task() calls that block until complete:

• Use when you need the result before proceeding
• Use for sequential dependencies
• Default mode — most common

4. Custom Agents (inline briefs)

Trigger: "custom agents", "spin up agents", "launch agents", "specialized agents" Action: Write a distinct brief for each agent — a name, a role/expertise, and a stance — then launch each with Task(subagent_type="general-purpose"). No composition tool; the orchestrator writes the persona.

Task(
  subagent_type="general-purpose",
  prompt=<agent brief: "You are a skeptical security reviewer focused on auth bypass..."> + <task + context>
)

• Give each agent a DIFFERENT brief so their perspectives actually diverge
• A capable model writes a sharper topic-specific persona than any trait lookup
• Ideal for: domain experts, adversarial reviewers, creative brainstormers, parallel analysis

5. Agent Teams (via TeamCreate)

Trigger: "create an agent team", "agent team", "swarm", "team of agents" Action: Use TeamCreate tool → TaskCreate → spawn teammates via Task(team_name=...) → coordinate via SendMessage

1. TeamCreate(team_name="my-project")           # Creates team + task list
2. TaskCreate(subject="Implement auth module")   # Create team tasks
3. Task(subagent_type="general-purpose", team_name="my-project", name="auth-engineer")  # Spawn teammate (senior-engineer/TDD brief in prompt)
4. TaskUpdate(taskId="1", owner="auth-engineer") # Assign task
5. SendMessage(type="message", recipient="auth-engineer", content="...")  # Coordinate

This is a COMPLETELY DIFFERENT system from custom agents:

• Custom agents (inline briefs) = fire-and-forget parallel workers, no shared state
• Agent teams (TeamCreate) = persistent coordinated teams with shared task lists, messaging, multi-turn

Team Guidelines:

• Use for 3+ independently workable criteria at Extended+
• Large complex coding tasks benefit most
• Each teammate works independently on assigned tasks via shared task list
• Parent coordinates via SendMessage, reconciles results
• Teammates go idle between turns — send messages to wake them

When to Use Teams vs Subagents (Decision Matrix)

Decision rule: If agents need to talk to each other or iterate on shared work → Teams. If each agent does independent one-shot work → Subagents.

Concrete examples:

• "Research 4 topics in parallel" → Subagents (independent, no coordination needed)
• "Build a feature spanning API + UI + tests with shared state" → Teams (cross-layer, needs coordination)
• "Run 10 file updates with same pattern" → Subagents (parallel, identical, independent)
• "Debug a complex issue with competing hypotheses" → Teams (need to share findings, adjust approach)

6. Parallel Task Dispatch

For N identical operations (e.g., updating 10 files with the same pattern):

1. Create N Task() calls in a single message (parallel launch)
2. Each agent gets one unit of work
3. Results collected when all complete

Effort-Level Scaling

Two-Tier Delegation (Lightweight vs Full)

Not all delegation needs a full agent. Match delegation weight to task complexity:

Lightweight Delegation

For: One-shot extraction, classification, summarization, simple Q&A against provided content.

Task(subagent_type="general-purpose", max_turns=3, prompt="...")

• Model is a per-dispatch judgment call (model param; unspecified inherits the session model)
• Set max_turns=3 — if it can't finish in 3 turns, it needs full delegation
• Provide all input inline in the prompt (no tool use expected)
• Examples: "Classify this text as X/Y/Z", "Extract the 5 key points from this", "Summarize this in 2 sentences"

Full Delegation

For: Multi-step reasoning, tasks requiring tool use (file reads, searches, web), tasks that need their own iteration loop.

Task(subagent_type="general-purpose", prompt="...")  # or specialized agent type

• Unspecified model inherits the session model (harness behavior)
• No max_turns restriction — agent iterates until done
• Agent uses tools autonomously (Read, Grep, Bash, etc.)
• Examples: "Research X and produce a report", "Refactor these 5 files", "Debug why test Y fails"

Decision Rule

Ask: "Can this be answered in one LLM call with no tool use?" → Lightweight. Otherwise → Full.

Why this matters: Spawning a full agent for a one-shot extraction wastes ~10-30s of startup overhead and unnecessary context. Lightweight delegation returns in 2-5s. Over an Extended+ Algorithm run with 10+ delegations, this saves minutes. Inspired by RLM's llm_query() vs rlm_query() two-tier pattern (Zhang/Kraska/Khattab 2025).

Right-Sizing Pre-Gate (PLAN, all tiers)

The tier delegation floors set a minimum fan-out. This gate sets the ceiling and the proof you owe. Run it before any fan-out. It exists because over-delegation is one of the most expensive recurring wastes in the reflection log: teams spawned for single-file rewrites, a writing agent that reported "completed" with zero disk writes (110k tokens spent for nothing), 300-agent waves with no headroom left to verify them. The outside proof is Cloudflare's risk-tiered dispatch — scale the agent count to the size of the job. Don't send the dream team to review a typo fix.

• (a) Zero-agent check. Is the answer already in working memory, or reachable by Glob+Grep+Read in under 30s, or isolated to a single file? Then 0 agents — do it inline. A subagent isn't free; its setup, context load, and result handling cost more than a direct read.
• (b) Disk-effect probe on every writing agent. An agent that says it wrote or edited files isn't trusted until you confirm it: the file exists AND the diff is non-empty (Read / git diff / Grep the claimed change). A "completed" report is a claim, not evidence. Rule 1 Live-Probe binds delegates exactly as it binds the primary.
• (c) Budget reservation above ~8 agents. A fan-out past ~8 concurrent agents must reserve explicit verification budget — you can't spend it all generating and none confirming — and name a non-agent fallback branch in ## Decisions for when the wave comes back unusable. This bounds the 5-level nesting capability: nesting multiplies agent count, so the ceiling is on the whole tree, not just the top layer.

Output at the Delegation Gate: 📐 RIGHT-SIZE: [0-agent inline | N agents, disk-probed | N>8, verify-budget reserved + fallback named].

Agent Briefs Are Ideal-State Prompts

Every inline brief and Workflow subagent prompt articulates the ideal state, not the procedure. State WHAT a done result looks like (as testable outcomes), the CONSTRAINTS, and the TOOLS available — then trust the agent to find HOW. A brief that choreographs the agent's reasoning ("first search, then extract, then verify, then synthesize") is BPE-violating scaffolding: it caps a capable agent and rots as models improve. This is the highest-volume prompting surface in the system (subagent briefs are written constantly at runtime), so the discipline matters most here. Keep the four keep-classes — safety-gate, verified-gotcha, tool-contract, output-format-contract — and delete the choreography. Prefer stating the coverage OUTCOME over a hardcoded fan-out count ("trust a match only when 3+ independent sources align", not "spawn exactly 15 agents"). Full standard: skills/Prompting/Standards.md § Ideal-State Prompting.

Anti-Patterns (Don't Do These)

• Don't delegate what Grep/Glob/Read can do in <2 seconds
• Don't spawn agents for single-file changes
• Don't create teams for fewer than 3 independent workstreams
• Don't send agents work without full context — they start fresh
• Don't use built-in agent names for custom agents
• Don't use bare built-in agent types when user asks for specialized or custom agents — write a distinct inline brief per agent and launch general-purpose
• Don't use full delegation for one-shot extraction/classification — use lightweight tier

Gotchas

• Delegation uses Claude Code's built-in TeamCreate for coordinated teams — distinct from one-shot custom agents (inline briefs on general-purpose). These are different patterns.
• 3+ independent workstreams warrant delegation. For 1-2 tasks, direct work is faster than team coordination overhead.
• Agent teams share a task list. Use TaskCreate/TaskUpdate for coordination, not ad-hoc messages.
• Teams overkill for single-file tasks. (Mar 2026 reflection: "one agent that can both read code and write JSX is better than three specialists who can't coordinate")
• Forked subagents inherit the full main-thread conversation + share its prompt cache (Anthropic CC v2.1.133+, enabled via CLAUDE_CODE_FORK_SUBAGENT=1 env or agent frontmatter context: fork). Use forked subagents for nuance-dependent work: design variations, follow-on research, anything that depends on context the main thread already established. DO NOT FORK for code review — a forked reviewer defends its own code (R Amjad: "biased toward defending own code"). For convergence questions, run one fork + one non-fork in parallel and look at where they disagree.

Examples

Example 1: Parallel implementation

User: "build the frontend and backend in parallel"
→ Creates team via TeamCreate
→ Spawns frontend and backend agents
→ Shared task list for coordination
→ Agents work independently, merge results

Example 2: Research swarm

User: "launch an agent team to research these 5 topics"
→ Creates team with 5 research agents
→ Each agent handles one topic independently
→ Results synthesized by team lead

Execution Log

After completing any workflow, append a single JSONL entry:

echo '{"ts":"'$(date -u +%Y-%m-%dT%H:%M:%SZ)'","skill":"Delegation","workflow":"WORKFLOW_USED","input":"8_WORD_SUMMARY","status":"ok|error","duration_s":SECONDS}' >> ~/.claude/LIFEOS/MEMORY/SKILLS/execution.jsonl

Replace WORKFLOW_USED with the workflow executed, 8_WORD_SUMMARY with a brief input description, and SECONDS with approximate wall-clock time. Log status: "error" if the workflow failed.