Research Pipeline

Build an autonomous research system that takes a topic, searches across web, academic papers, and your internal knowledge base, and delivers a structured report with source citations — reviewed by an editor agent for accuracy. Turn hours of research into minutes.

What you'll have

A research agent that searches multiple sources in parallel
An editor agent that validates claims and flags unsupported statements
A pipeline that chains research → editing in one API call
Knowledge graph entries that accumulate across research sessions

Step 1: Configure agents

yaml

# astra.yml — research pipeline
agents:
  - id: research-agent
    systemPromptTemplate: |
      You are a senior research analyst. When given a topic, conduct
      thorough multi-source research using web search, academic papers,
      and existing knowledge. Cite all sources. Produce structured reports.
    providers: [claude, gemini]
    tools:
      allow: [brave_search, arxiv_search, web_scrape, summarize, knowledge_retrieve]
    budget:
      maxCostCents: 300
      maxToolCalls: 50
      maxDuration: 300000   # 5 minutes

  - id: editor-agent
    systemPromptTemplate: |
      You are a technical editor. Review research reports for accuracy,
      clarity, and completeness. Flag unsupported claims.
    providers: [claude]
    tools:
      allow: [brave_search, summarize]

Step 2: Run the pipeline

bash

# Run a research pipeline: research → edit → publish
curl -X POST http://localhost:3000/pipelines/run \
  -H "Authorization: Bearer ${JWT_TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Analyze the current state of edge AI inference: hardware trends, latency benchmarks, and deployment patterns for 2026",
    "stages": [
      { "agentId": "research-agent" },
      { "agentId": "editor-agent", "systemPromptSuffix": "Focus on factual accuracy and source quality" }
    ]
  }'

Step 3: Seed with domain knowledge

Drop a context file into ./workspace/ to guide research focus and source preferences. The file watcher picks it up immediately.

markdown

# Drop domain knowledge into workspace files
# workspace/RESEARCH-CONTEXT.md
Research focus areas:
- Edge computing and IoT
- LLM inference optimization
- Semiconductor supply chain

Preferred sources:
- ArXiv, IEEE, ACM
- Company technical blogs (not marketing)
- Benchmark datasets from MLPerf

Going deeper

For multi-day research projects, use Deep Research mode which spawns sub-agent swarms for parallel investigation and delivers structured reports with progress tracking.

Research findings automatically flow into the knowledge graph, so each research session builds on the last. After a few runs, the agent starts connecting dots across topics that a human would miss.