Skip to content

IDP Agent

The IDP Agent is the main agent that searches and analyzes documents and generates artifacts through conversation with users. It operates using Strands SDK’s ReAct pattern, combining MCP tools and Code Interpreter for complex tasks.

User question
AgentCore Runtime (HTTP streaming)
Strands Agent (Claude Opus 4.6)
├─ 1. Understand intent
├─ 2. Create execution plan
├─ 3. Load skill → call tools → collect results
└─ 4. Generate final response with citations

The agent operates in skill units. Skills are markdown files defined in .skills/{name}/SKILL.md that the agent reads and follows before performing tasks.

SkillPurposeTools Used
searchDocument search + web search strategySearch MCP, Graph MCP, DuckDuckGo
docxWord document creation/editingCode Interpreter (python-docx)
xlsxExcel spreadsheet creation/editingCode Interpreter (openpyxl)
pptxPowerPoint creation/editingCode Interpreter (python-pptx)
diagramStructural diagram generationCode Interpreter (Mermaid)
chartData visualization chart generationCode Interpreter (Matplotlib)
qa-analysisQA analysis managementQA MCP
markdownMarkdown document generationMD MCP
User: "Summarize the V-101 valve analysis results in Word"
├─ [1] Load search skill → document search
│ ├─ Search MCP (summarize) → vector + FTS search
│ └─ Graph MCP (graph_search) → entity connection traversal
├─ [2] Load docx skill → Word document creation
│ └─ Code Interpreter → write document with python-docx → S3 upload
└─ [3] Final response with citations
→ [document_id:doc_xxxxx](s3_uri)
→ [artifact_id:art_xxxxx](filename.docx)

MCP tools accessed through the AgentCore Gateway.

ToolDescription
summarizeHybrid search (vector + FTS) → Haiku summarization, returns qa_ids
overviewList project documents
ToolDescription
graph_searchEntity graph traversal based on qa_ids, discovers related pages
link_documentsCreate document-to-document links
unlink_documentsRemove document-to-document links
ToolDescription
extract_textExtract text from PDF/DOCX/PPTX
extract_tablesExtract tables from documents
create_documentCreate PDF/DOCX/PPTX
edit_documentEdit existing documents
MCPToolDescription
Image MCPanalyze_imageImage analysis
QA MCPget_document_segmentsRetrieve document segments
QA MCPadd_document_qaAdd QA analysis
MD MCPload_markdownLoad markdown
MD MCPsave_markdownSave markdown
MD MCPedit_markdownEdit markdown

AgentCore Code Interpreter provides an isolated Python sandbox environment. AWS SDK is pre-configured, enabling S3 uploads.

Used when the agent generates artifacts (documents, charts, diagrams).

Code Interpreter
├─ python-docx, openpyxl, python-pptx (document generation)
├─ matplotlib (charts)
├─ mermaid-py (diagrams)
└─ boto3 (S3 upload)
→ s3://{bucket}/{user_id}/{project_id}/artifacts/{artifact_id}/

The project’s language setting is fetched from DynamoDB and injected into the system prompt. The agent responds in that language.