Cloud Research

Inspiration

The next bottleneck in research may not be a shortage of intelligence. It may be human bandwidth.

AI's potential in research is no longer speculative. Progress in mathematical reasoning, scientific discovery, coding, and literature analysis suggests that machine researchers will soon be able to explore more hypotheses, read more papers, and run more intellectual experiments than any individual human could follow. That creates a new problem: can people still understand, challenge, and take intellectual ownership of research produced at machine speed?

Most auto-research projects ask: How much of the scientist can we remove from the loop? That frames the human as latency. We reject that premise. Research is not complete when a system emits an answer or a paper. It becomes useful when a person can understand it, contest it, reproduce it, and decide what it means. The more consequential question is: How much more capable can a scientist become when AI works around them?

A researcher should not have to personally push every intermediate step forward. Searching the latest work, checking code, comparing mechanisms, running a cheap counterexample, recording a dead end, and turning scattered evidence into the next decisive question all consume the same scarce resource: human attention. When the researcher sleeps, teaches, studies, or focuses on another problem, those processes usually stop.

Cloud Research began with a different vision: give every researcher a persistent research institution of their own. The human remains the PI—the source of purpose, taste, judgment, and responsibility. A panel of AI researchers maintains momentum around that human, and an Explainer makes every important result re-enterable. The goal is not research without people. The goal is to let people think at the scale of the research world they now inhabit.

What it does

The researcher creates a Lab, defines the experts it needs, and gives it one brief: the field, the current question, prior attempts, constraints, compute budget, and what a useful handoff should contain. A Gemini Research Director reads that brief and recruits specialists in whatever order the evidence demands.

A typical Lab contains:

Finder, which searches current papers, code, and nearby ideas for the real opening rather than a keyword-level gap.
Theorist, which turns an opening into a precise mechanism, prediction, proof, or counterexample.
Experimentalist, which runs the cheapest decisive probe before recommending expensive compute.
Critic, which attacks novelty, evidence, assumptions, and alternative explanations, then proposes the strongest repair.
Writer, which preserves sources, claims, failed paths, artifacts, and the next executable research task.
Explainer, which reconstructs a paper, task, experiment, or full research shift so the human understands not only what the Lab concluded, but why its beliefs changed.

These are defaults, not a fixed organization chart. The human can create multiple Labs; add, remove, rename, and rewrite roles; address the whole panel; or @ one expert directly. The Director can call experts repeatedly and in any order. There is no scripted scientific state machine pretending to know the shape of a discovery in advance.

When the researcher leaves, the same Lab continues as a Cloud Run Job. It can search with live Google grounding, compare claims, run isolated Python probes, challenge weak directions, preserve failures, and produce an evidence-backed Handoff. Firestore records the Lab, run, expert events, artifacts, and final Handoff across scale-to-zero and browser departure. A signed GitHub Issue label can launch a shift without an open tab, and Writer can prepare the research packet on a dedicated GitHub branch.

When the human returns, Cloud Research does not present a magical final answer. It presents a research inheritance:

what the Lab investigated;
which candidates survived and which were killed;
what evidence changed the argument;
what remains unknown;
which cheap experiments were run;
what deserves the researcher's attention next; and
how to continue the work locally or in the cloud.

The researcher can then question the Explainer in text or speak with the animated Lab Bot through Gemini Live. Human participation happens where it matters—setting direction, interrogating evidence, changing the team, and making scientific judgments—not through an endless sequence of approval buttons.

How we built it

Cloud Research uses Gemini 3.7 Flash through Vertex AI, Google ADK 2.7.1, Cloud Run, and Cloud Firestore. Gemini Live provides the native-audio debrief with the central Lab Bot.

At the center is one ADK Director. Every custom specialist is exposed to it as an ADK AgentTool. The Director sees the brief, the current evidence, and the available experts, then decides whom to call next. Google Search grounding is available to research roles. Experimental roles receive Gemini's isolated code execution. Writing roles receive artifact and GitHub tools. Explainer can be called inside a research shift or used independently for a paper, task, result, or previous Handoff.

The cloud architecture separates interaction from durable execution:

A private Cloud Run service hosts the Lab interface and accepts authenticated briefs.
A Cloud Run Job owns the long research shift, so closing the browser does not terminate it.
Firestore acts as an append-only fact ledger for Labs, expert events, delivery records, artifacts, and Handoffs.
A minimal public Cloud Run gateway verifies HMAC-signed GitHub webhooks, rejects duplicate deliveries, selects the requested Lab, and launches the Job.
Secret Manager and separate service identities keep the public event surface isolated from the private research surface.

We deliberately kept the scientific runtime thin. Cloud Run manages lifetime, retries, and scale-to-zero; Firestore remembers facts; ADK exposes specialists; Gemini decides the scientific move. Our prompts carry three pieces of research taste:

Concise: give the expert a sharp responsibility, not a policy manual.
Excited: pursue the crack that could change understanding; research should remain deeply interesting.
Certain: trust that valuable research fruit exists, even when the current candidate is wrong.

The shared operating belief is: Trust the opportunity. Doubt the candidate. Keep going.

Cloud Research was built from scratch during the hackathon submission period. The application, research architecture, prompts, custom roles, interface, Cloud Run surfaces, Firestore ledger, GitHub gateway, Live experience, and deployment were created specifically for this project.

Challenges we ran into

The first challenge was resisting the temptation to encode research as a workflow diagram. A conventional multi-agent system wants stages, gates, statuses, and deterministic routing. Real research is not known in advance. A source can invalidate the premise; a critic can expose a missing variable; a four-line probe can make an entire direction irrelevant. Hard-coding the path would make the demo look orderly while making the Lab less intelligent.

We solved this by putting scientific judgment in short expert prompts and letting the ADK Director choose the path. The code manages capabilities and persistence, not scientific taste.

The second challenge was making autonomy legible. A system that performs many actions while the human is away can easily become alienating: the user returns to conclusions they did not witness and cannot reconstruct. We therefore treated explanation as a first-class research capability, not a summary generated at the end. Expert messages, failed hypotheses, sources, experiments, and belief changes remain available to the Explainer and to the human PI.

The third challenge was continuity. A browser request is short-lived; a serious research shift is not. We had to carry a custom Lab definition and run identity into a background Job, stream real expert events back into the interface, survive refreshes and cold starts, and reconstruct the same Handoff from Firestore when the researcher returned.

The fourth challenge was evidence honesty. Grounded models can still produce weak associations or bad links. In our source-transport audit, 23 of 24 unique URLs resolved and one generated GitHub URL returned 404. We preserved that failure instead of editing it away. Reachability is not scientific support; it is only one auditable layer in the evidence chain. The Critic and human PI still need to judge whether a source actually supports a claim.

Finally, we had to make a sophisticated system feel calm. Multiple Labs, custom experts, live events, voice, background execution, and durable artifacts could easily become a dashboard full of status machinery. We designed the interface as a research room: one conversation, a visible team, a fluid Lab Bot, and a Handoff that teaches rather than overwhelms.

Accomplishments that we're proud of

We built a genuine model-directed expert panel whose order is selected by evidence, not by a disguised workflow graph.
A research shift survives the browser as a Cloud Run Job and scales back to zero when finished.
Researchers can create multiple Labs and completely rewrite the number, identity, and role of their experts.
A real signed GitHub event can launch the selected Gemini Lab, preserve its events in Firestore, and create a research branch without an open browser.
Explainer supports papers, tasks, results, and the Lab's own work, making comprehension part of the product rather than an afterthought.
Gemini Live provides an interruption-ready voice debrief grounded in the current Handoff, with the Lab Bot visibly listening, thinking, and speaking.
Three cross-domain research Jobs completed with seven to eight real expert events each, including killed paths, unknowns, source lists, artifacts, and next actions.
A separate bounded Cloud Run proof path completed five consecutive executions with one succeeded task and exit code 0 in every run.
The public surface is reduced to a health endpoint and a verified GitHub gateway; the research service, Lab data, model calls, and Live sessions remain private.

What we are proudest of is not that the system can produce more research text. It is that the Lab can work independently without pushing the human out of the intellectual process.

What we learned

We learned that the real unit of research is not a task. It is a continuing argument between ideas, evidence, experiments, critics, and people. A useful research agent therefore needs continuity and memory more than a longer answer box.

We learned that autonomy without legibility is a weak form of augmentation. If the human cannot inherit the reasoning, uncertainty, and next question, the system has accelerated output while slowing understanding. Explanation is not presentation polish; it is the interface through which agency returns to the researcher.

We learned that multi-agent value does not come from showing many avatars. It comes from productive differences: one expert searches, another commits to a mechanism, another tries to destroy it, and another preserves what changed. The disagreement is the feature.

We also learned that optimism and rigor are allies. Believing that an important opening exists gives the Lab permission to reject weak candidates without becoming timid. A failed hypothesis is useful when it is preserved as information and converted into a sharper next move.

Most importantly, we learned that the best role for AI in research is not simply replacing human labor. It is expanding human scientific agency: giving one person more eyes, more critics, more working memory, and more uninterrupted time—while keeping meaning and responsibility with the person.

What's next for Cloud Research

The next step is to turn one research shift into a living research program.

We want persistent claim lineage across weeks: which source introduced a belief, which experiment changed it, which critic challenged it, and which later result repaired or rejected it. We want every Handoff to connect to the next Lab shift instead of becoming another forgotten report.

We will add secure compute bridges that let a researcher attach their own CPU or GPU server. Cloud agents will be able to prepare a bounded experiment, hand it to authorized compute, observe the result, and bring the evidence back into the same research conversation. Heavy compute remains a resource owned by the human; the Lab makes it easier to use deliberately.

We also want teams of humans to participate in the same research institution: multiple PIs, student–advisor handoffs, domain-specific Labs, replication panels, and an Explainer that adapts to what each collaborator already understands. Claim-level source audits and reproducible experiment lineage will make the system increasingly suitable for serious scientific work.

The long-term vision is not a machine that publishes while people watch. It is a new interface between human curiosity and machine-scale exploration—a place where AI keeps research moving and humans can keep understanding, questioning, and deciding what the research means.