Skip to content

PATTERN Cited by 1 source

Flattened tool architecture

Intent

Collapse every product's capabilities from separate sub-agent tool surfaces into a single, uniformly named, directly-callable tool surface for the orchestrator LLM. Nothing is paraphrased through a second agent.

Mechanism

Every operation is exposed as a typed, namespaced action: jira__search_issues, confluence__get_page, google_calendar__list_events. The orchestrator's LLM calls these directly and sees raw arguments, raw responses, and raw errors — not a paraphrase from an intermediate agent.

When a call fails or returns something unexpected, the same LLM that made the call reads the failure and decides what to do next: retry with different arguments, fall back to a related tool, or surface the gap to the user.

Trade-offs

Benefit Cost
End-to-end visibility of tool results + errors Large flat tool catalog risks selection degradation
Recovery is part of the reasoning loop Mitigated by patterns/progressive-tool-disclosure-meta-tools
One model migration (not N sub-agents) Requires careful namespace design
No per-product schema tax per invocation Loses sub-agent prompt specialisation (mitigated by SKILL.md)

Complementary patterns

  • Progressive disclosure — avoids paying schema cost for the full flattened surface upfront.
  • SKILL.md — replaces the per-product prompt expertise that previously lived inside sub-agent system prompts.

Seen in

Last updated · 542 distilled / 1,571 read