Structured Intermediate Representations

Definition

Standardizing heterogeneous subagent outputs into a common structured intermediate (a claim/evidence/source/confidence format) via a conversion layer before a synthesis agent consumes them.

Key points

• The loss: passing raw text from financial and news agents to a synthesis agent results in tables losing clarity and news losing narrative flow.
• Flow: Financial Agent (Structured JSON), News Agent (Prose Summaries), Patent Agent (Structured Lists) → Format Conversion Layer (standardizes outputs to a common intermediate representation) → JSON (claim, evidence, source, confidence) → Synthesis Agent (Executive Briefings).
• Citation rule: to prevent lost attributions, require all subagents to output structured claim-source mappings that the synthesis agent is instructed to preserve.
• On the Architect's Reference Matrix this is the Accuracy × Multi-Agent cell ("Structured Intermediates").
• Pairs with Goal-oriented delegation and Shared Memory Architecture in the multi-agent stack.

Why it matters for the exam

• Multi-agent synthesis scenarios: the fix for lossy hand-off is a common structured intermediate + preserved claim-source mappings, not free-text concatenation.

Common gotchas

• Concatenating raw agent text is the anti-pattern; convert to a common schema first.
• Attributions must be preserved end-to-end via claim-source mappings.

See also

• Goal-oriented delegation
• Shared Memory Architecture
• Architect's Reference Matrix

Sources

• architects-playbook
• netlight-study-guide