Deep Research Agent

A specialized agent for conducting comprehensive web research on any given topic. The agent provides thorough, accurate, and well-structured research reports through systematic web searches, analysis, and synthesis.

Features

• Systematic Research Workflow: Structured approach from topic analysis to final report generation
• Parallel Search Capabilities: Efficient concurrent web searches for multiple queries
• Comprehensive Context Management: Tracks research progress, queries, and findings
• Automatic Report Generation: Creates markdown research reports with bash tools
• Rate Limit Handling: Built-in tools to manage API rate limits gracefully
• Error Recovery: Robust error handling and retry mechanisms

Core Tools

Research Management Tools

• updateResearchTopicTool: Set the main research topic
• updateRationaleTool: Define research approach rationale
• updateQueriesTool: Generate targeted search queries
• updateQueriesDoneTool: Track completed searches
• updateKnowledgeGapTool: Identify information gaps
• updateIsSufficientTool: Mark research completeness
• updateSummariesTool: Consolidate findings

Execution Tools

• WebSearchTool: Conduct web searches with multiple sources
• BashCommandTool: Execute bash commands for report generation
• WaitingTool: Handle rate limits and API throttling
• TodoUpdateTool: Manage task lists and progress tracking
• AgentStopTool: Signal completion of research

Rate Limit Management

The Deep Research Agent includes built-in rate limit handling to prevent API throttling issues:

WaitingTool Usage

When encountering rate limit errors (e.g., OpenAI token limits), the agent can:

// Example usage
WaitingTool({
  seconds: 20,
  reason: "Rate limit encountered, waiting before retry"
})

Error Recovery Strategy

1. Detection: Automatically recognizes rate limit error patterns
2. Waiting: Uses WaitingTool to pause execution (15-30 seconds)
3. Retry: Resumes the failed operation after waiting
4. Escalation: Reports persistent failures to user

Rate Limit Best Practices

• Wait 15-30 seconds for OpenAI rate limits
• Use specific reasons for waiting (tracking and debugging)
• Continue with exact same action after wait period
• Monitor token usage through enhanced context management

Research Workflow

1. Topic Analysis Phase

• Set research topic with context
• Define research approach rationale
• Generate diverse, targeted search queries

2. Web Search Phase

• Execute parallel searches for efficiency
• Track query completion status
• Gather information from multiple authoritative sources

3. Synthesis Phase

• Consolidate findings into comprehensive summaries
• Organize information logically with citations
• Maintain objectivity and balanced perspectives

4. Reflection Phase

• Evaluate research completeness
• Identify knowledge gaps
• Generate additional queries if needed

5. Completion Phase

• Generate comprehensive markdown research report
• Save report to research.md using BashCommandTool
• Inform user of report location
• Signal completion with AgentStopTool

Example Usage

import { DeepResearchAgent } from '@continue-reasoning/deep-research';
import { OPENAI_MODELS } from '@continue-reasoning/core';

const agent = new DeepResearchAgent(
  'research-agent',
  'Deep Research Specialist',
  'Conducts comprehensive web research',
  30, // max steps
);

// Start research
await agent.startWithUserInput("Research the latest developments in quantum computing");

Configuration

Enhanced Prompt Processing

The agent uses enhanced prompt processing with optimized message retention:

• Messages: 100 steps
• Tool calls: 20 steps (increased for complex research)
• Errors: 5 steps
• Thinking: 5 steps
• Reasoning: 3 steps

Context Management

Research context automatically tracks:

• Research topic and rationale
• Pending and completed queries
• Knowledge gaps identification
• Research sufficiency status
• Consolidated summaries
• Task progress (todos)

Error Handling

Common Scenarios

• Rate Limits: Automatic waiting and retry with WaitingTool
• API Throttling: Intelligent backoff strategies
• Search Failures: Alternative query generation
• Network Issues: Graceful degradation and retry

Monitoring and Debugging

• Comprehensive logging of all operations
• Tool execution tracking
• Error pattern recognition
• Performance optimization

Output

The agent generates:

1. Real-time Progress: Step-by-step research updates
2. Comprehensive Report: Final markdown document with:
   • Executive summary
   • Key findings and insights
   • Source citations and references
   • Balanced analysis and conclusions
3. Structured Data: Persistent context for future reference

Testing

Run the waiting tool test to verify functionality:

npx tsx packages/deep-research/examples/test-waiting-tool.ts

This ensures the rate limit handling mechanism works correctly before conducting research.