Consensus rebuilt its science search as a four-agent workflow on GPT-5 and the Responses API

What the front end hides: four agents doing one researcher's job

Consensus started as what founder Christian Salem calls a vertical search engine for science: index papers, retrieve results, summarize with citations. The new capability, Scholar Agent, keeps that same conversational surface but restructures everything underneath into four agents with distinct jobs.

A Planning Agent decomposes the question and decides the next action. A Search Agent works the paper index, a user's private library, and the citation graph. A Reading Agent interprets papers individually or in batches. An Analysis Agent synthesizes results, chooses structure and visuals, and composes the output. The user still types a question and gets an answer — the coordination is invisible.

By dividing the workflow across agents, we reduce error and make the system far more disciplined. No one agent has too much responsibility, which turns out to be key for reliability.Montana Labs

The 'research context pack' is a trust affordance, not a backend detail

Consensus describes its method as context engineering — assembling the right evidence before generation begins. What matters for the user experience is what that produces: every answer ships with a structured bundle of papers, metadata, and key findings that trace back to original studies.

The narrow scope of each agent is framed as a way to minimize hallucinations, and the architecture explicitly allows the assistant to decline. If no studies clear its quality threshold, it says so rather than fabricating an answer. For a tool used by researchers and, increasingly, clinicians, the ability to return nothing is a feature the interface has to make legible.

We don't want researchers wasting time double-checking every claim. If the system can't ground an answer in real evidence, it won't make one up.Montana Labs

Why the Responses API migration mattered for the routing

Consensus moved from Chat Completions to the Responses API specifically to support multi-agent routing, citing better reliability and cost efficiency and finer control over sub-agent calls. That's the plumbing behind a four-agent handoff — each planning, search, reading, and analysis step is a separate model call that has to be orchestrated cleanly.

The team also reports that in early evaluations GPT-5 outperformed GPT-4.1, Sonnet 4, and Gemini 2.5 Pro on tool-calling accuracy and planning stability. Their stated conclusion is telling: reliable tool-calling let them spend less time on prompt gymnastics and more on agent behaviors that map to how research actually works.

A consumer front end is what made a research back end scale

The strategic bet is that the interface, not the institution, is the distribution channel. Consensus went direct-to-researcher — students, faculty, clinicians — and built something that feels like a modern consumer app: fast onboarding, intuitive design, conversational interface. Adoption spread by word of mouth from PhD candidates to faculty to doctors.

That surface is now attached to consequential users. Consensus signed the Mayo Clinic's medical library and launched a 'Medical Mode' for practitioners searching clinical evidence, alongside 8 million users and 8x revenue growth. The implication for teams building agentic products: the multi-agent workflow, the context pack, and the refusal-to-answer discipline only earn their keep if the front end stays simple enough that a clinician trusts it on the first try — the engineering exists to protect a consumer-grade experience, not the other way around.