Prompting multi-agent orchestration: delegation rules and failure modes

How should a lead AI agent plan, brief and synthesise the work of parallel worker agents in a multi-agent orchestration system? AR Labs researched this question in July 2026, and the answer shaped our Orchestrate and multi-agent research skills. The report draws on Anthropic's account of its multi-agent research system and on community analyses of why multi-agent LLM systems fail, and condenses them into delegation rules and a failure-mode taxonomy.

Anthropic-official findings (high confidence)

From "How we built our multi-agent research system" (Anthropic engineering, 2025), paraphrased:

• Canonical loop: orchestrator plans, decomposes and synthesises; workers execute isolated slices and return only conclusions; workers never talk to each other.
• Extended thinking as pre-delegation scratchpad: the orchestrator reasons through complexity, fleet size, division of labour, and each delegation prompt before any spawn.
• Two-level parallelism (90% latency win): lead spawns 3–5 workers in parallel; each worker batches 3+ tool calls in parallel. The second level is the one everyone forgets.
• Effort calibration: simple fact-find = 1 agent, 3–10 calls; comparison = 2–4 agents; complex research = 10+ with clear divisions.
• Four mandatory delegation fields: objective, output format, tool guidance, task boundary. Boundary is the most-omitted; require workers to describe their seams with neighbouring agents.
• CitationAgent pattern: a dedicated verifier checks final-report claims against source documents, not the lead agent's recollection.
• Rubric judges: five criteria (factual accuracy, citation accuracy, completeness, source quality, tool efficiency). Un-rubric'd judges preferred SEO content over academic PDFs.
• Poor tool descriptions sent agents down wrong paths; a tool-inventory pass cut completion time 40%.

Failure-mode taxonomy (community, well-sourced)

1. Runaway fleets: retry blips in recursive systems multiply calls 27× (3×3×3). Fix: hard caps in prompt, and surface failures as findings; never retry indefinitely.
2. Worker duplication: fix with boundaries that name what neighbours own.
3. Concatenation instead of synthesis: fix with an explicit synthesis pass (conflicts, then integrate, then attribute).
4. Telephone effect: evidence degrades across hops; fix with raw-evidence fields (URL, verbatim quote) in worker schemas.
5. Orchestrator does work itself then delegates the remainder, contaminating briefs.
6. Cold-start thrash (Claude Code): workers see CLAUDE.md but not the conversation; brief them "like the stranger it is": paths, errors, decisions, schema.
7. Permission-gate deadlock (Claude Code): background workers auto-deny permission prompts; scope read-only tasks to read-only agents.
8. Hallucination conformity: same-model verifiers conform to confident claims; use fresh-context refute-framed verification, cross-model where possible.
9. Context rot: attention degrades mid-context before the window fills, compounding hop by hop; workers return compact conclusions, and files (not conversation) carry handoffs.

Not covered

Agent SDK 2026 orchestration primitives; benchmarked topology comparisons (hub-and-spoke vs pipeline vs mesh); token-accounting patterns for fleet budgeting.

Sources

• Anthropic engineering: How we built our multi-agent research system (read via ai.plainenglish.io mirror; theaiengineer.substack.com deep-dive; zenml.io LLMOps summary, mutually consistent)
• Redis: Why multi-agent LLM systems fail, and Cemri et al., Why Do Multi-Agent LLM Systems Fail? (arXiv 2503.13657)
• hidekazu-konishi.com: Claude Code subagents & orchestration guide (strongest Claude-Code-specific source)
• digitalapplied.com: Multi-Agent Orchestration: 5 Patterns That Work in 2026
• fountaincity.tech: Anthropic's Multi-Agent Blueprint: What Production Adds
• ivern.ai: Claude Code Subagents guide (2026)