Agent-assisted mode
Agent-Assisted analysis means describing what you need in natural language. The Agent Workbench ships with a built-in Codex runtime, and the Agent is not a generic chatbot — it is wired to the platform through MCP: it queries tools, searches databases, reads task logs, and submits actions under a confirmation gate.
Typical flow
Natural-language question → MCP tool discovery → database / resource queries → governed submission → result interpretation
For every request the Agent first performs capability routing, sorting it into one of four categories: direct execution / adapt-then-execute / needs more input / needs exploration — runnable jobs go straight ahead, missing inputs are requested first, and exploratory needs get a candidate route that advances step by step.
Enter the Agent Workbench
- Open “Agent Workbench” from the platform home page navigation;
- The workbench session is bound to your user and project, and the backend writes the MCP configuration automatically;
- Describe your need as if chatting, and upload or reference files already on the platform when needed.
[Screenshot coming: Agent Workbench main interface]
What the Agent can do
Capabilities the Agent reaches through MCP (see MCP governance):
- Query tool schemas: know which tools exist and how to fill their parameters, so tool selection is grounded;
- Search databases and workspace files: gather evidence across databases and reference workspace and task-artifact files;
- Read task logs and result manifests: diagnose failures and interpret output files;
- Submit governed actions: write operations such as submitting analysis tasks, executed only after your confirmation.
What the confirmation gate feels like
- Read-only operations (L1, 48 tools) run immediately: tool lookups, database queries, and log reads never interrupt you;
- Write operations (L2, 7 tools) confirm first: every time the Agent is about to submit a task or run a batch job, it shows the exact call and parameters in the conversation and waits for your click before submitting;
- The workspace file tool (1 tool) is constrained by workspace boundaries and session scope: e.g. updating an R script requires more explicit authorization.
* Tier counts are the 2026-10-16 static catalog count (56 expanded tools: 48 read, 7 controlled write, 1 workspace file); actual per-session visibility is per tools/list.
You can always decline, or ask the Agent to revise its plan before confirming.
[Screenshot coming: a confirmation interaction for an L2 write operation]
Platform skills: pre-installed domain playbooks
The Agent Workbench ships with 9 platform skills. A skill is a structured playbook written for the Agent (when to use it, which steps to follow, what output to produce). You never invoke them manually — the Agent matches and triggers them automatically based on your request:
| Skill | What it does | When it triggers |
|---|---|---|
| platform-capability-planner | Capability discovery and tool selection; lists candidates only, never submits | You describe an analysis goal but haven’t settled on a tool |
| mcp-task-lifecycle | Read live schema → assemble parameters → submit task → save the task_id | You’ve decided which platform tool to run |
| task-results-reader | Read-only access to task status, logs, and artifact lists | You have a task_id and want progress, results, or failure causes |
| bulk-rnaseq-analysis | Bulk RNA-seq project orchestration: quantification, DE, enrichment, reports, resumption | Re-runs, resumptions, and batch orchestration of RNA-seq projects |
| gene-identity-gate | Gene identity verification gate; prevents ID/alias mix-ups | When ID interpretation affects conclusions or cross-database merging |
| rstudio-analysis-runtime | Controlled execution and maintenance of R scripts in RStudio containers | You explicitly ask to run or modify R scripts in RStudio |
| academic-paper-polish | Unified routing for paper polishing (auto-splits mixed Chinese/English drafts) | You ask to polish a paper without specifying a language, or the draft is mixed |
| chinese-academic-polish | Conservative Chinese academic polishing; audits before rewriting | You explicitly ask to polish Chinese text |
| plant-paper-writing | English paper polishing with gene nomenclature rules and ESL checks | You explicitly ask to polish English text |
When writing analysis requests, state the goal, species, data, and thresholds clearly — the Agent will pick the right skill and follow its playbook.
A typical conversation
User: Here is a maize candidate-gene table; assemble the transcription factor
regulatory evidence for each gene.
Agent:
1. Parse the uploaded gene table and verify gene identities
2. gene_id_mapper unifies gene IDs
3. Query PlantTFDB / JASPAR / PlantRegMap
4. motif_to_tf_mapping links motifs to TFs
5. Summarize everything into a candidate TF evidence table
Every tool call enters the MCP audit and can be reviewed in the audit records afterwards.
MCP access channels
The platform provides two equivalent channels for external MCP clients: compat (/api/mcp/*) and official (/api/mcp-official/*). See MCP governance.
Security boundaries
- The Agent never runs arbitrary shell; it only calls MCP-exposed tools;
- L1 reads need no confirmation, L2 writes need confirmation, Workspace file operations follow session scope;
- Large results are returned through the Result Store with pagination to avoid context overflow;
- Every call records audit metadata: user, permission scope, confirmation state, and more.