> [!WARNING]
> **Preview**
>
> These features are Generative AI Preview offerings, subject to the "Pre-GA Offerings Terms" of the [Google Cloud Service Specific Terms](https://cloud.google.com/terms/service-terms). For these Generative AI Preview offerings, Customers may elect to use them for production or commercial purposes, or disclose Generated Output to third-parties, and may process personal data as outlined in the [Cloud Data Processing Addendum](https://cloud.google.com/terms/data-processing-addendum), subject to the obligations and restrictions described in the agreement under which you access Google Cloud.
>
> The AI/ML Data Location terms also apply to Co-Scientist, as an Agentic AI Service on Gemini Enterprise, subject to its restrictions and those listed in the [Google Cloud Platform Services with Data Residency](https://cloud.google.com/terms/data-residency) page.

Co-Scientist is a Google-developed agent that accelerates research and
development processes. It uses Gemini to automate complex scientific
reasoning. For algorithmic discovery, see the
[AlphaEvolve documentation](https://docs.cloud.google.com/gemini/enterprise/docs/alphaevolve/developer-guide/overview.md.txt).

> [!NOTE]
> **Note:** Access to Co-Scientist is restricted. To request access, contact your Google account team. Your account team provides detailed product documentation once they grant access.

## Overview of Co-Scientist agent

Co-Scientist is a multi-agent AI system built on Gemini, designed to act as a
virtual scientific collaborator. The system helps researchers uncover original
knowledge and formulate novel research hypotheses and experimental proposals.
By combining AI reasoning with the scientific method, Co-Scientist enables
scientists to navigate literature, synthesize insights across domains, and
accelerate biomedical discoveries.

![Co-scientist](https://docs.cloud.google.com/static/gemini/enterprise/docs/images/coscientist-alphaevolve/coscientist.png)

### How it works

Co-Scientist is built on a multi-agent architecture and an asynchronous task execution
framework to scale test-time computation. This design allows the system to
reason, debate, and improve outputs iteratively.

When you provide a research goal in natural language, Co-Scientist coordinates
the following specialized agents:

- **Generation agent:** Explores literature using web search, synthesizes findings, and formulates initial hypotheses.
- **Reflection agent:** Acts as a peer reviewer to examine the correctness, quality, and safety of generated hypotheses.
- **Ranking agent:** Evaluates and prioritizes research proposals using an Elo-based tournament.
- **Evolution agent:** Refines top-ranked hypotheses by grounding them in literature and combining innovative ideas.
- **Proximity agent:** Organizes tournament matches and deduplicates hypotheses using a proximity graph.
- **Meta-review agent:** Synthesizes insights from all reviews and debates into a comprehensive research roadmap.

A dedicated **Supervisor agent** orchestrates the process, allocates resources,
and manages the worker queue to maintain a self-improving loop of scientific
reasoning.

### Use cases

Co-Scientist solves complex problems across multiple scientific disciplines,
specifically those requiring deep subject matter expertise and
trans-disciplinary insights. Example use cases include:

- **Drug repurposing:** Identify novel therapeutic indications for approved
  drugs by analyzing molecular signatures, signaling pathways, and
  literature-based information.

- **Target discovery:** Propose new biological components and epigenetic targets
  for complex diseases and streamline hypothesis selection for experimental
  validation.

- **Mechanistic explanations:** Generate hypotheses for intricate problems, such
  as antimicrobial resistance, by synthesizing fragmented scientific data.

For more information and additional use cases, see the
[Accelerating scientific breakthroughs with an AI Co-Scientist](https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/) blog.

## Monitor Co-Scientist runs with Cloud Logging

When a Co-Scientist run completes successfully, Gemini Enterprise
writes one log entry to Cloud Logging in your Google Cloud project. The entry
records who started the run, when it started, how long it took, and how many
tokens it used. You can use these entries to track Co-Scientist usage per
user, attribute token consumption to teams, or build dashboards and alerts.

> [!NOTE]
> **Note:** An entry is written only for runs that complete successfully. Runs that fail, are paused, or are deleted before they finish don't produce an entry.

### Before you begin

- Turn on observability for the Gemini Enterprise app that runs Co-Scientist. In the app, go to **Configurations** \> **Observability** and turn on **Enable instrumentation of OpenTelemetry traces and logs** . The page says these settings configure the Core Assistant agent, but Co-Scientist run entries also follow them. Settings on the Co-Scientist agent's own page don't affect these entries. You don't need to turn on **Enable logging of prompt
  inputs and response outputs** . For more about these settings, see [Manage observability settings](https://docs.cloud.google.com/gemini/enterprise/docs/manage-observability-settings.md.txt).
- To view the logs, you need the [Logs Viewer](https://docs.cloud.google.com/iam/docs/roles-permissions/logging.md.txt#logging.viewer) IAM role (`roles/logging.viewer`) on the Google Cloud project.

Entries are written only for runs that complete after observability is turned
on. Earlier runs aren't logged retroactively.

### View run entries

1. In the Google Cloud console, go to the **Logs Explorer** page.

   [Go to Logs Explorer](https://console.cloud.google.com/logs/query)
2. Select the Google Cloud project that contains your Gemini Enterprise
   app.

3. Enter the following query and click **Run Query**:

   ```sql
   logName="projects/PROJECT_ID/logs/discoveryengine.googleapis.com%2Fgemini_enterprise_user_activity"
   jsonPayload.logMetadata.methodName="CoscientistRunCompletion"
   ```

   Replace the following:
   - `PROJECT_ID`: the ID of your project.

To use the Google Cloud CLI instead, run the following command:

```bash
gcloud logging read \
  'logName="projects/PROJECT_ID/logs/discoveryengine.googleapis.com%2Fgemini_enterprise_user_activity" AND jsonPayload.logMetadata.methodName="CoscientistRunCompletion"' \
  --project=PROJECT_ID \
  --freshness=7d \
  --format='table(timestamp, jsonPayload.logMetadata.name.basename(), jsonPayload.coscientistRunInfo.starterUserPrincipal, jsonPayload.coscientistRunInfo.inputTokenCount, jsonPayload.coscientistRunInfo.outputTokenCount, jsonPayload.coscientistRunInfo.duration)'
```

To narrow the results, add conditions to the query. For example:

- Runs started by one user: `jsonPayload.coscientistRunInfo.starterUserPrincipal="alex@example.com"`
- Runs for one app: `jsonPayload.logMetadata.name:"/engines/APP_ID/"`
- Runs in one location: `resource.labels.location="us"`

Each entry's `jsonPayload.coscientistRunInfo` records who started the run, when
the instance was created, how long it ran, and how many input and output tokens
it used. `jsonPayload.logMetadata.name` identifies the Co-Scientist instance.
For the full list of fields, see
[Logged information](https://docs.cloud.google.com/gemini/enterprise/docs/set-up-usage-audit-logs.md.txt#logged-information).
To control who can read these entries, see
[Control access to logs](https://docs.cloud.google.com/gemini/enterprise/docs/set-up-usage-audit-logs.md.txt#control_access_to_logs).