IDP Agent

Overview

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 4.8, default)
  ├─ 1. Understand intent
  ├─ 2. Create execution plan
  ├─ 3. Load skill → call tools → collect results
  └─ 4. Generate final response with citations

Model Selection

Users can pick the model and reasoning effort per turn from the chat composer.

The selectable models are loaded at runtime from a catalog defined in an SSM parameter (/idp-v2/chat/models). Adding/removing a model is done by editing the parameter — no redeploy needed.
The agent validates the requested model_id against the catalog allowlist (the frontend selector is not a security boundary). A model that is not allowed falls back to the default.
The reasoning effort (low/medium/high) is passed as output_config.effort only when the resolved model supports it.
Changing the model starts a new conversation, so responses from different models don’t mix within one session.

Skill System

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.

Skill	Purpose	Tools Used
search	Document search + web search strategy	Search MCP (summarize, graph_traverse, graph_keyword), AgentCore Web Search
dataset	Structured data (Excel/CSV) Text2SQL querying	Data MCP (search_datasets, describe_dataset, run_sql)
docx	Word document creation/editing	Code Interpreter (python-docx)
xlsx	Excel spreadsheet creation/editing	Code Interpreter (openpyxl)
pptx	PowerPoint creation/editing	Code Interpreter (python-pptx)
diagram	Structural diagram generation	Code Interpreter (Mermaid)
chart	Data visualization chart generation	Code Interpreter (Matplotlib)
qa-analysis	QA analysis management	QA MCP
markdown	Markdown document generation	MD MCP

Execution Flow

User: "Summarize the V-101 valve analysis results in Word"
  │
  ├─ [1] Load search skill → document search
  │   ├─ Search MCP (summarize) → vector + FTS search
  │   └─ Search MCP (graph_traverse) → 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

MCP tools accessed through the AgentCore Gateway.

Search MCP

Tool	Description
`summarize`	Hybrid search (vector + FTS) → Haiku summarization, returns qa_ids
`graph_traverse`	Entity graph traversal based on qa_ids, discovers related pages
`graph_keyword`	Keyword similarity search via LanceDB graph keywords + Neptune traversal
`overview`	List project documents

Document MCP

Tool	Description
`extract_text`	Extract text from PDF/DOCX/PPTX
`extract_tables`	Extract tables from documents
`create_document`	Create PDF/DOCX/PPTX
`edit_document`	Edit existing documents

Data MCP (Structured Data Text2SQL)

Queries Parquet datasets converted from Excel/CSV using SQL. Used for structured data that needs exact aggregation, filtering, or sorting.

Tool	Description
`search_datasets`	Hybrid search over the project’s datasets by name/description
`describe_dataset`	Retrieve a dataset’s schema, samples, and query examples (reference doc)
`run_sql`	Run read-only SQL against a Parquet dataset via DuckDB

Other MCPs

MCP	Tool	Description
Image MCP	`analyze_image`	Image analysis
QA MCP	`get_document_segments`	Retrieve document segments
QA MCP	`add_document_qa`	Add QA analysis
MD MCP	`load_markdown`	Load markdown
MD MCP	`save_markdown`	Save markdown
MD MCP	`edit_markdown`	Edit markdown

Local Tools

Tools that run directly in the agent process, not through the Gateway.

Tool	Description
`render_chart`	Render an inline chart card in the reply (hbar / compare / timeline / donut / stacked / scatter). Uses only real values from a prior tool result
`ask_user`	Show a structured question card (single / multi / free-text). The user’s choice is posted back as the next message and picked up on the following turn
`generate_image`	AI image generation
`code_interpreter`	Isolated Python sandbox execution (see below)

Results from render_chart and ask_user are rendered by the frontend as dedicated UI cards instead of plain text.

Code Interpreter

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}/

Multi-language Support

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