PATTERN Cited by 1 source

Single-loop agent orchestration¶

Intent¶

Replace a hierarchical multi-agent system (coordinator + specialist sub-agents) with one LLM, one context, one iterative loop. The single model holds the full conversation state, calls tools directly, reads raw responses and errors, and makes all recovery decisions within the same reasoning context.

Motivation¶

Hierarchical multi-agent systems create lossy hand-offs: the coordinator sees only summaries from sub-agents, never raw tool outputs, intermediate reasoning, or error details. The single-loop approach keeps everything in one context so:

End-to-end error recovery happens in the same context that planned the call.
No information is compressed or dropped between agents.
Model upgrades are a single workstream (one model, not N sub-agents to re-tune).
The model can reason across products without reconstructing what a sub-agent meant.

Shape¶

User query → Build prompt → Loop (up to N iterations):
  │
  ├── Call LLM with full context + available tools
  │     ↓
  ├── Tool call(s)? → Execute (parallel if independent)
  │     → Add results to context → Continue loop
  │
  └── Final answer? → Deliver to user with citations

Trade-offs¶

Benefit	Cost
End-to-end visibility and recovery	Context window pressure grows with iterations
Single model migration	Requires context compaction to stay within limits
Cross-product reasoning	Larger tool surface needs progressive disclosure
Simpler architecture	Higher per-iteration latency than quick sub-agent dispatch

When it fits¶

Frontier models with long context. Models that handle 100k+ tokens and large tool surfaces reliably.
Tasks requiring cross-domain reasoning. Information from one product informs decisions about another.
Need for error recovery. The same model that fails can retry, fall back, or explain the gap.

When it doesn't fit¶

Ultra-low-latency simple lookups. A quick routing dispatch to a specialised lightweight agent may be faster.
Extremely wide tasks. Decompose into child instances (variant of patterns/context-segregated-sub-agents) rather than one loop.

Seen in¶

sources/2026-06-18-atlassian-long-horizon-reasoning-engine — Atlassian Long Horizon: single LLM, single context, up to 150 iterations. Replaced Hybrid Orchestrator (coordinator + per-product sub-agents). +8.5% offline accuracy, +23% task completion.