BioResearchAgent

Architecture comparison

Agentic complexity is not assumed to help — it's measured

Three architectures, same evidence budget, same synthesis prompt. Differences in outcome are attributable to the architecture, not to prompt drift between them.

Baseline 0 · no retrieval

A single LLM call, no retrieval at all. The control condition every other architecture must beat.

latency: 74.4s
tokens: 10124
failures: 0/20

Baseline 1 · retrieve + synthesize

BM25-retrieve on the raw question, then one LLM call to synthesize an answer from the retrieved passages.

latency: 126.9s
tokens: 51493
failures: 0/20

Agent 1 · planner + retrieve + synthesize

An LLM call first decomposes the question into 1-3 targeted search queries, then retrieves and synthesizes exactly as Baseline 1 does — isolating the planner's effect.

latency: 160.8s
tokens: 54212
failures: 1/20

Baseline 1 vs. Agent 1, fully and fairly graded

MetricBaseline 1 (20/20)Agent 1, worst-case (20/20)Winner
Unsupported-claim rate0%6%Baseline 1
Claim support rate100%94%Baseline 1
Citation precision100%95%Baseline 1
Source validity rate100%95%Baseline 1
Abstention correct95%90%Baseline 1
Retrieval recall@k85%78%Baseline 1
Concept coverage40%45%Agent 1
Retrieval MRR57%65%Agent 1

On task B4, Agent 1’s planner split a vague question into two sharper sub-queries and recovered a complete retrieval miss that Baseline 1 suffered (recall@k 0.0 → 1.0) — a genuine, mechanistic win for query planning.

On task D1, Agent 1’s own planner call returned syntactically invalid JSON (a real model error, not a parsing bug) and the task failed outright, while Baseline 1 answered it correctly.

Charging that one failure the worst-case value on every metric it would otherwise report — rather than letting it silently drop out of the denominator — reverses the naive read: Agent 1 loses on 6 of 8 shared metrics and wins only on concept coverage and retrieval ranking quality (MRR), at roughly 27% more latency. This number went through three honest revisions as two separate grading bugs were found and fixed; see methods for the full account.