Skip to main content
The Discovery Agent identifies scientifically rigorous discoveries from completed tasks and links them to supporting evidence. It maintains traceability from claims to tasks to job IDs.

​
Function Signature

// src/agents/discovery/index.ts
export async function discoveryAgent(input: {
  conversationState: ConversationState;
  message: Message;
  tasksToConsider: PlanTask[];
  hypothesis?: string;
}): Promise<DiscoveryAgentResult>;

type DiscoveryAgentResult = {
  discoveries: Discovery[];
  start: string;
  end: string;
};

type Discovery = {
  claim: string;
  evidence: string;
  novelty: string;
  taskId: string;      // Links to PlanTask
  jobId?: string;      // External job ID for traceability
  artifactIds?: string[]; // Generated artifacts supporting this claim
};

​
What It Does

​
Identifies Novel Claims

Extracts scientifically rigorous discoveries from ANALYSIS tasks: 
const result = await discoveryAgent({
  conversationState,
  message,
  tasksToConsider: [
    {
      id: "task-1",
      type: "ANALYSIS",
      objective: "Analyze differential gene expression",
      output: "Analysis identified 342 DEGs (FDR < 0.05). Key regulators: TP53 (log2FC=-2.3), MYC (log2FC=1.8).",
      jobId: "bio-analysis-123",
      artifacts: [
        { id: "artifact-1", filename: "volcano_plot.png" }
      ]
    }
  ],
  hypothesis: "p53-p21 axis regulates cellular senescence"
});

// Returns:
// {
//   discoveries: [
//     {
//       claim: "TP53 shows significant downregulation (log2 fold change -2.3, FDR < 0.05) in senescent cells",
//       evidence: "Differential expression analysis of RNA-seq data",
//       novelty: "Confirms p53 involvement in senescence pathway",
//       taskId: "task-1",
//       jobId: "bio-analysis-123",
//       artifactIds: ["artifact-1"]
//     }
//   ]
// }

​
Updates Existing Discoveries

Enriches existing discoveries with new evidence: 
conversationState.values.discoveries = [
  {
    claim: "p53 regulates senescence",
    evidence: "Literature review",
    taskId: "task-0"
  }
];

const result = await discoveryAgent({
  conversationState,
  message,
  tasksToConsider: [newAnalysisTask],
  hypothesis
});

// Merges or extends existing discoveries

​
Maintains Traceability

Each discovery links back to:
  • taskId: Which task produced this finding
  • jobId: External job ID (Edison/BioAgents)
  • artifactIds: Generated files (plots, tables)
This enables complete traceability in paper generation.

​
Discovery Structure

​
claim
string
required
The scientific claim or finding
​
evidence
string
required
Supporting evidence from analysis
​
novelty
string
required
What makes this discovery novel or important
​
taskId
string
required
ID of task that produced this discovery
​
jobId
string
External job ID for full traceability
​
artifactIds
string[]
IDs of generated artifacts (plots, tables, code)

​
Usage Example

// src/routes/deep-research/start.ts
const discoveryResult = await discoveryAgent({
  conversationState,
  message,
  tasksToConsider: completedMaxTasks,
  hypothesis: hypothesisText
});

// Update conversation state
conversationState.values.discoveries = discoveryResult.discoveries;

await updateConversationState(
  conversationState.id,
  conversationState.values
);

​
Incremental vs Full Updates

The caller decides which tasks to analyze:
Pass only new tasks to update discoveries:
const newTasks = completedMaxTasks.filter(t => 
  !previousTaskIds.includes(t.id)
);

await discoveryAgent({
  conversationState,
  message,
  tasksToConsider: newTasks,
  hypothesis
});

​
Discovery Criteria

The agent identifies discoveries that are:
Claims must be supported by data analysis, not just literature synthesis.
Includes specific metrics (fold changes, p-values, effect sizes).
Either new findings or validation of hypothesis.
Linked to specific task outputs and job IDs.

​
Integration with Paper Generation

Discoveries flow into paper generation with full traceability: 
// src/routes/deep-research/paper.ts
const discoveries = conversationState.values.discoveries || [];

for (const discovery of discoveries) {
  // Link to task output
  const task = completedTasks.find(t => t.id === discovery.taskId);
  
  // Link to artifacts
  const artifacts = task?.artifacts?.filter(a => 
    discovery.artifactIds?.includes(a.id)
  );
  
  // Include in paper
  paperContent += `\n\nDiscovery: ${discovery.claim}\n`;
  paperContent += `Evidence: ${discovery.evidence}\n`;
  paperContent += `Artifacts: ${artifacts.map(a => a.filename).join(", ")}\n`;
}

​
Configuration

​
STRUCTURED_LLM_PROVIDER
string
default:"openai"
LLM provider: openai, anthropic, google, or openrouter
​
STRUCTURED_LLM_MODEL
string
default:"gpt-5"
Model name for discovery extraction

​
Historical Note

Previously, discovery extraction was handled by the Reflection Agent. It’s now a separate agent for better separation of concerns:
  • Reflection Agent: Updates world state (objective, insights, methodology)
  • Discovery Agent: Identifies and structures scientific claims

Reflection Agent

Updates world state with insights

Paper Generation

Uses discoveries for LaTeX papers

Analysis Agent

Generates analysis outputs

Build docs developers (and LLMs) love

Get started for freeTalk to us