Agent Studio helps you understand how the AI curation agents work and gives you tools to improve them. You can browse agent prompts, build visual curation workflows, and chat with Claude Opus about any of it.
Click "Agent Studio" in the navigation bar at the top of the application.
Agent Studio has three main tabs: Agents, Flows, and Agent Workshop. All tabs include a chat panel where you can talk with Claude Opus.
On both tabs, the left panel is your chat with Claude Opus. You can ask Opus about whatever you're viewing on the right - agent prompts or flows.
Browse the instructions given to each AI agent and chat with Opus about them.
Agent Browser (Right Panel)
See all agent prompts organized by subcategory:
- System - Supervisor Agent that routes your queries to specialists (internal, not available in Flow Builder)
- PDF Extraction - PDF Extraction Agent and Gene Expression Extractor
- Data Validation - Gene, Allele, Disease, Chemical, GO Term, GO Annotations, Ortholog, and Ontology Mapping agents
- Output - Chat Output, CSV Formatter, TSV Formatter, JSON Formatter agents
- My Custom Agents - Custom agents you have created in Agent Workshop
- Shared Agents - Custom agents shared by other users in your project
The Agent Browser also includes filter tabs (All, Shared, Templates) at the top of the agent list to help narrow down agents quickly.
For each agent, you can view:
- Base Prompt - The core instructions given to the agent
- Group-Specific Rules - How the prompt is customized for each curator group (WormBase, FlyBase, MGI, ZFIN, RGD, SGD, Xenbase)
- Combined View - See the base prompt with group rules injected
- Tools - The tools available to each agent (listed in the agent card)
Clickable Tool Names
Tool names in agent cards are clickable! Click any tool name to open a detailed panel showing:
- Description - What the tool does
- Parameters - Input parameters with types and descriptions
- Methods - For multi-method tools (like database queries), see all available methods with examples
- Agent Context - Which methods are relevant to the selected agent
This helps you understand exactly what capabilities each agent has and how they interact with databases and APIs.
Ask Opus about agents:
- "Why does this agent look for negative evidence?"
- "I think this prompt is missing guidance about [organism-specific convention]"
- "Can you help me write a suggestion to improve this?"
- "What does this instruction mean in practice?"
Build visual curation workflows and chat with Opus about them. See Curation Flows for the complete guide to building flows.
Flow Builder (Right Panel)
Create workflows by dragging agents onto a canvas and connecting them:
- 15 available agents from extraction to file output
- Save, load, and reuse flows
- Generate downloadable CSV, TSV, or JSON files
Verify with Claude (Important!)
Before running a flow, click the "Verify with Claude" button. Claude will:
- Check your flow structure for issues
- Identify missing connections or problematic configurations
- Suggest improvements
- Confirm your flow is ready to run
This is especially valuable when building new flows or troubleshooting ones that aren't working as expected.
Ask Opus about flows:
- "Does this flow make sense for extracting expression data?"
- "What agent should I add to map anatomy terms to WBbt IDs?"
- "Why isn't my flow generating the output I expected?"
Create and test custom versions of agent prompts without affecting the live system.
What is a Custom Agent?
A custom agent is your personal copy of a system agent's prompt. You can edit the instructions, add per-group overrides, and use it in flows — all without changing anything for other users. Custom agents you create also appear in the Flow Builder agent palette under "My Custom Agents".
Getting Started
- Choose a Getting Started mode - Select Template, Scratch, or Clone
- Set your base configuration - Pick template/clone source (if applicable), model, and tools
- Edit the prompt - Modify instructions and optional per-group overrides
- Choose an icon - Pick an emoji icon from the icon picker to identify your agent
- Save - Use File → Save Agent. Each save creates a version you can revert to later.
You can also get here quickly from the Agents tab: click "Clone to Workshop" on any agent's detail panel.
File Menu
The Workshop toolbar provides these operations:
- New Agent - Start a new custom agent draft
- Open Agent... - Search and open a previously saved custom agent
- Manage Agents... - Open or delete saved custom agents
- Save Agent / Save New Agent - Save your current work (creates a new version)
- Delete Agent - Remove the current custom agent
Icon Picker
When creating or editing a custom agent, select an emoji icon to help identify your agent in the palette and flow canvas. Available icons include 🔧, 🧬, 📄, 🔍, 🧪, 📊, 🧠, ⚙️, ✨, 📝, 📚, 🧩, and more.
Per-Group Prompt Overrides
Customize how your agent behaves for different curator groups:
- Check "Include group rules at runtime" to apply group-specific rules when your custom agent runs
- Select a group from the dropdown (e.g., WormBase, FlyBase, MGI)
- Edit the override - Write group-specific instructions in the text area
- Reset to Template - Revert a group override back to the template version
The Workshop shows which groups have custom overrides and a total override count.
Version History
Every save creates a new version. You can add optional save notes to describe your changes. Revert to any previous version if an edit doesn't work out.
Template Availability
If a template source is unavailable for an older custom agent, Agent Workshop will show a warning that the custom agent cannot be executed until a valid template is selected.
Tool Library
Attach tools to your custom agent from the available tool library:
- In the Advanced Settings section, expand the Tools accordion
- Click "Manage Tools" to open the Tool Library dialog
- Browse or search tools by name or category (Database, API, Document, Output)
- Check the tools you want to attach and click "Done"
Some tools may be marked as "Not attachable by policy" -- these are reserved for specific system agents.
Tool Requests
If you need a new tool that does not exist yet:
- Click "Need a new tool? Ask Claude" to draft a request with Claude's help
- When ready, click "Send to Developers" to submit a formal tool request
- Track the status of your requests in the Tool Requests accordion (pending, reviewed, in progress, completed, or declined)
Visibility
Control who can see your custom agent:
- Private - Only you can see and use this agent (default)
- Shared with Project - Other users in your project can see and use this agent in their flows
Model Selection
Choose the AI model that powers your custom agent:
- Select a model from the Model dropdown (the Workshop shows model descriptions and recommendations)
- If the model supports reasoning levels, choose a level (low, medium, high) -- higher levels are slower but better for difficult tasks
- Not sure which model to pick? Click "Confused about models? Chat with Claude" for guidance
Using Custom Agents in Flows
Custom agents appear in the Flow Builder palette under "My Custom Agents". You can drag them into flows just like system agents.
Discuss with Claude
Click the "Discuss with Claude" button in the Workshop toolbar to send your current draft prompt to Opus for review. When the Agent Workshop tab is active, the left-panel chat is aware of your workshop context — your selected template source, draft prompt, and group settings. You can ask Claude to:
- "Critique this draft and suggest concrete edits"
- "Help me restructure this prompt for clarity"
- "What would happen if I changed this instruction?"
If you want to talk about the results from a conversation you're having in the main chat, you can bring that into Agent Studio:
- In the main chat, find the AI response you want to discuss
- Click the triple-dot menu (⋮) on that message
- Select "Open in Agent Studio"
This opens Agent Studio with your conversation loaded, so Opus knows exactly what you're referring to. You can then ask questions like:
- "Why did the AI suggest this ontology term instead of that one?"
- "The AI missed the gene mentioned in paragraph 3 - what went wrong?"
- "Can you help me understand why I got this response?"
This is the best way to get help understanding unexpected AI behavior or to formulate improvement suggestions.
Your domain expertise is invaluable for improving the AI. Agent Studio provides two ways to submit suggestions:
Click the "AI-Assisted" button in the chat header. Opus will:
- Review your conversation
- Summarize the issue you've discussed
- Draft a concrete improvement suggestion
- Submit it for your confirmation
When to use: After discussing a specific issue with Opus, this is the fastest way to submit actionable feedback.
Click the "Manual" button to fill out a suggestion form yourself:
- Suggestion Type:
- Improvement - General enhancement
- Bug - Incorrect or unexpected behavior
- Clarification - Ambiguous instructions
- Group-Specific - Change needed for your group
- Missing Case - Scenario the prompt doesn't handle
- General - Feedback not tied to a specific prompt
- Summary - Brief description (1-2 sentences)
- Detailed Reasoning - Why this change is needed
- Proposed Change (optional) - Specific wording you'd suggest
When to use: When you have a clear suggestion and don't need Opus's help drafting it.
Instead of "The AI is wrong," try "When I asked about gene X, the AI said Y, but according to [source], it should be Z."
If you see a pattern of errors, describe 2-3 specific cases. This helps identify whether it's a prompt issue or something else.
If your group has specific naming conventions, annotation rules, or curation practices that the AI doesn't follow, explain them. You're the expert on your organism!
Before suggesting a change to a base prompt, check if your group already has specific rules in the Agent Browser. The issue might be that your group's rules need updating rather than the base prompt.
When providing feedback about a specific interaction, always use the triple-dot menu to open it in Agent Studio rather than describing it manually. This gives Opus (and the development team) the full context.
When you submit a suggestion:
- It's sent to the development team with full context
- The team reviews it for feasibility and impact
- If approved, the prompt is updated in the next deployment
- Complex suggestions may require discussion before implementation
Your suggestions help make the AI better for everyone!
No. If you have feedback based on a conversation or general observation, you can submit "General" feedback without selecting a specific agent.
Yes! That's the main purpose of the Agents tab. Browse all agent prompts and see exactly what instructions each agent receives.
Each curator group has organism-specific conventions and curation practices. Group rules customize the AI to respect these conventions, for example using WormBase anatomy terms (WBbt) for C. elegans or FlyBase allele naming patterns.
Each agent has a base prompt, optional group-specific rules, and optional flow custom instructions. These combine in a defined priority order: flow custom instructions (highest) > base prompt > group rules. Flow instructions override everything else for that step. See How Prompts Layer Together in the Curation Flows guide for full details.
The main chat uses a multi-agent system optimized for curation tasks - it routes your questions to specialists who query databases. Agent Studio's Opus is for discussing how the AI works, understanding specific responses, and improving the system.
See the Curation Flows guide for complete documentation.
If you have questions about using Agent Studio or need help formulating feedback, just ask Opus! It's designed to help you translate your domain expertise into actionable suggestions.