Qwestly Agentic Orchestration — Research Docs

A collection of research, planning, and architectural thinking for building Qwestly's agentic orchestration layer.

Docs Index

Doc	What it covers
Architecture Overview	Foundational concepts: agent topologies, orchestration patterns, tool-use, RAG, memory. Required reading before evaluating frameworks.
Framework Comparison	Head-to-head comparison of the top agentic frameworks in 2026 — LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, Pydantic AI, Semantic Kernel, Dify, and more. Trade-offs, sweet spots, Python vs TS.
RAG & Tools Patterns	How to wire up retrieval-augmented generation and tool-calling into an agent system. Chunking strategies, embedding pipelines, hybrid search, function-calling schemas, MCP protocol.
Qwestly Orchestration Plan	The concrete proposal. A proposed architecture for Qwestly's single-entry-point chatbot that routes intent to sub-agents (LinkedIn ingestion, Qwestly Card generation, profile Q&A, etc.).
Testing Agentic Systems	How testing changes when your code calls an LLM. Unit tests, integration tests, E2E tests, evals (LLM-as-judge), CI/CD strategy. Anti-patterns and framework-specific patterns.
Deployment Options	Can you run an agent orchestrator on Vercel? Timeouts, memory, cold starts, streaming, and deployment patterns. Given your existing FastAPI on Vercel setup.
Unstructured Memory	NEW — LLM-managed agent memory. How Qwestly remembers user preferences, goals, facts, and decisions across conversations. Extraction, storage, retrieval, consolidation, and decay. Three-tier data architecture (structured / memory / RAG).
Human-in-the-Loop	How to gate dangerous actions (publishing to LinkedIn, deleting data) behind human approval. Tool-level vs orchestrator-level vs UI-level gating. Framework support across Pydantic AI, LangGraph, OpenAI SDK.
Bonus Topics	Prompt engineering for agents (system prompt structure, tool descriptions, chain-of-thought), cost management & model tiering, security & prompt injection, production observability, multi-turn conversation management.

How to read this

If you're new to agentic systems: start with Architecture Overview, then Framework Comparison, then the rest.

If you already know the landscape and want the takeaway: jump to Qwestly Orchestration Plan and use the other docs as reference.

Status

These docs were generated as part of an exploratory / research phase. Nothing here is set in stone — the goal is to equip you to make informed decisions about stack, architecture, and implementation order.