E8 · Agent teams on a task queue

Principle

A coordinator launches persistent agents. Each one takes its tasks from a shared queue and builds up domain context on its own scope as the work goes on.

The problem it solves

On sustained, multi-step work, starting a fresh agent for every task makes it lose what it learned about its scope.

When to use it

• Sustained, multi-step work, where each teammate gets familiar with its scope.
• Tasks are strongly independent of one another.

When to avoid it

• Tasks depend on one another: the pattern needs strong independence, and sharing intermediate results is hard. Prefer E1.
• Detecting when the work is done is tricky: without a clear stopping criterion, the system can spin idle.

Common pitfalls

The anti-patterns that misuse this pattern, best established first. Each sheet describes the symptoms, the public incidents and the remedies.

Loop, stopping and orchestration

• AP40 · Parallel writers — Several agents edit the same deliverable at once, each with its own implicit decisions.
• AP43 · No durable execution or idempotence — A failure halfway through, and the agent starts over, sending again the email it had already sent.
• AP44 · Big-bang agent deploy — A new version replaces the old one in one go, right in the middle of the agent sessions that relied on it.

Security and permissions

• AP65 · Agent that rewrites its own configuration — The agent can write to the files that set its permissions: an injection no longer bypasses approval, it switches it off.

All anti-patterns

Synonyms by vendor

• Agent teams (Anthropic (2026))

Neighbouring patterns

• E1 · Orchestrator and subagents
• E9 · Message bus
• E10 · Shared state (blackboard)

Sources

• Multi-agent coordination patterns: five approaches and when to use them — Anthropic (C. Phillips), 2026-04-10
• Why Do Multi-Agent LLM Systems Fail? — Cemri et al., arXiv, 2025-03