What it does AgentKit was BCG X’s full-stack starter kit for building “constrained” agent apps—chatbots that follow pre-defined action plans rather than wandering off into ReAct rabbit holes. It wired together a Next.js frontend, FastAPI backend, LangChain orchestration, and Postgres/pgvector in Docker Compose, with Celery/Redis for queues and NextAuth for authentication.

The interesting bit The core idea was reliability through restriction: a Meta Agent picks from human-authored Action Plans, then executes tool chains in parallel with streaming intermediate results to the UI. Think of it as guardrails for LLMs that still let you show off—tables, visualizations, code snippets—while the agent plods through its script.

Key highlights

• Pre-built Docker Compose setup with environment variable templates
• Streaming UI with rendering for tables, code blocks, and agent status updates
• Parallel tool execution within linear action steps
• Optional add-ons: user feedback collection, per-user settings, GitHub/email auth
• LangSmith integration for evaluation

Caveats

• Deprecated and unmaintained as of the README; BCG explicitly recommends Vercel AI SDK or FastAPI’s official full-stack template instead
• README warns it is “a starter kit,” not production-ready, and omits enterprise security features
• Documentation site (agentkit.infra.x.bcg.com) may or may not persist; the repo itself is archival

Verdict Worth a quick skim if you’re studying how consultancies packaged agent patterns circa 2023–2024, or if you need to maintain a legacy prototype. Start something new here and you’re building on abandoned scaffolding.