Here is the technical truth: real AI agent stacks are software systems. They are not prompt collections. They are not automations with a nicer sales page. They need architecture.
What sits inside the stack
We use LangGraph for orchestration when a workflow needs state, branching, and multi-step decisions. We use FastAPI to expose clean endpoints and keep integrations predictable. We use PostgreSQL plus pgvector when the system needs structured data and retrieval in one place. We use Redis for caching, queues, and speed.
LiteLLM helps us route models cleanly. LangSmith helps us trace, inspect, and debug what the agents are doing. Docker keeps environments stable. GitHub Actions handles deployment flow and repeatability.
Why model choice matters
We use different Claude models for different jobs. Haiku is fast and cheap for lighter routing or formatting tasks. Sonnet is the workhorse for everyday operational reasoning. Opus is for heavier judgment, deeper analysis, and higher-stakes outputs where quality matters more than speed.
What the team actually looks like
This work needs a lead developer, an automation engineer, and an operator who knows the business cold. One person can write code. That does not mean one person can design a useful system for a contractor operation.
We also think a lot about A2A protocol and Agent Cards because agents need clear roles, permissions, and handoffs. If everything is one giant fuzzy assistant, nothing is accountable.
Five questions to ask any AI company
- How do you orchestrate multi-step workflows when the agent needs memory and branching?
- What do you use for tracing, monitoring, and debugging?
- How do you choose between models and control cost?
- Where does company data live, and how is retrieval handled?
- Who on your team has actually run the kind of operation you are building for?
If the answers are vague, you are probably looking at a template reseller. Real systems sound specific because they are specific.