Two Papers Cut Agent Architecture Orthodoxy: One Strong Model Beats Pipelines, and Grep Beats Embeddings on Search
Two papers published this week challenge the two most common defaults in production agent engineering: that more agents improve reasoning, and that vector retrieval is the right foundation for agentic search. Both conclusions are narrower than the headlines suggest — and more useful for that reason.
Paper 1: Single-Agent Wins on Multi-Hop Reasoning
“Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets” (arXiv 2604.02460, Stanford)
The finding is almost embarrassingly clean. When thinking-token budgets are matched, a single agent reasoning internally outperforms sequential, debate, role-based, and ensemble multi-agent setups on FRAMES and MuSiQue — the standard multi-hop reasoning benchmarks. The evaluation covers Qwen, DeepSeek, and Gemini.
The mechanism is not surprising in hindsight. Multi-hop reasoning requires maintaining a chain of intermediate conclusions across multiple steps. A single agent keeps that chain in its internal context. A multi-agent system has to serialise it into messages. Every message boundary is a compression step. The Data Processing Inequality puts a formal ceiling on how much information can survive compression; in practice, the ceiling is lower than it looks.
Multi-agent systems only pulled ahead in one condition: when the single agent’s context was deliberately degraded — via masking, substitution, or misleading distractors. When the reasoning environment is hostile or unreliable, external coordination can partially compensate for corrupted internal state. That is a useful boundary condition, not a general endorsement.
The practical read: for multi-hop tasks (research, planning, legal analysis), start with one strong model. Reach for multi-agent pipelines when context corruption is the specific failure mode — not as a default architectural upgrade.
Paper 2 (and 3): Grep Beats Embeddings for Agentic Search
“Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction” (arXiv 2605.05242)
“Is Grep All You Need? How Agent Harnesses Reshape Agentic Search” (arXiv 2605.15184)
Two papers, same week, same conclusion: for agents doing evidence-location work, terminal search outperforms semantic retrieval — and costs less.
The DCI paper’s headline number: swapping semantic retrieval for Direct Corpus Interaction raised BrowseComp-Plus accuracy from 69% to 80% while lowering cost. The grep paper found inline grep beats inline vector search across every harness-model pair on LongMemEval tasks.
Neither paper argues that vector databases are useless. The argument is narrower: semantic retrieval is optimised for paraphrase tolerance, which is a liability when the answer is an exact string — a function name, a file path, a date, an error message. Embeddings pull in semantically adjacent documents that look close to the query but contaminate the model’s working context with near-misses. Grep retrieves nothing it did not find exactly.
For coding agents especially, a large fraction of real work is evidence-location: find the symbol, trace the call, inspect the diff. The DCI paper frames this as the agent having “fingers” rather than a lookup service — direct access to the corpus lets the model search an exact string, notice a related entity, constrain the next search, and iterate.
The cost differential is explained by the same mechanism: fewer tokens of spurious context passed to the model per query.
The Combined Implication
Both results point in the same direction: many of the assumed architectural upgrades in agent design — more agents, smarter indexes — were providing gains that had more to do with added compute or retrieval volume than fundamental capability improvements. Matching the budget and measuring carefully changes the comparison.
Neither paper says to abandon multi-agent systems or vector search. The more useful claim is that the default choice is wrong: start simple, measure on your specific task type, and treat added complexity as a hypothesis to test rather than a free upgrade.
All three papers are available on arXiv with evaluation code and datasets.