I need to build an agent that can read docs, query a database, and create tickets, what stack should I use?

A solid stack for an agent that reads docs, queries a database, and creates tickets is usually:

Recommended stack

1. Agent/orchestration layer

• Python for the main agent runtime
• LangGraph if you want reliable multi-step workflows and tool use
• OpenAI API or another LLM provider for reasoning and tool calling

Why:

• Python has the best ecosystem for data access, parsing, and automation.
• LangGraph is better than a simple prompt loop when the agent needs state, retries, and approvals.

---

2. Document ingestion / retrieval

• Unstructured, LlamaIndex, or LangChain loaders to ingest docs
• Postgres + pgvector if you want a simpler, production-friendly vector store
• Or Pinecone / Weaviate / Qdrant if you want managed or dedicated vector DBs

For most teams:

• Start with Postgres + pgvector
• Use LlamaIndex or LangChain for retrieval if you want faster implementation

---

3. Database access

• Postgres is the default recommendation if you can choose
• Use:
  • SQLAlchemy for DB access in Python
  • Or direct SQL for tighter control
• If the agent needs to query business data safely:
  • expose a read-only DB user
  • consider a SQL validation layer before execution

---

4. Ticket creation / integrations

Depends on your ticketing system:

• Jira: Jira REST API
• Linear: Linear API
• ServiceNow: ServiceNow REST API
• Zendesk: Zendesk API
• Or use Zapier/Make only for low-complexity workflows

For production, call the APIs directly from your agent service rather than relying on no-code tools.

---

5. App/API layer

• FastAPI for exposing the agent as a service
• Add:
  • auth
  • request logging
  • rate limiting
  • audit trails

---

6. Background jobs / reliability

• Celery + Redis or RQ for async tasks
• Or use Temporal if workflows are long-running and need strong reliability

If the agent will:

• ingest docs
• run DB queries
• create tickets asynchronously
  then a job queue is very useful.

---

7. Observability and guardrails

• LangSmith or OpenTelemetry for tracing
• Pydantic for structured outputs and validation
• Human approval step for ticket creation if actions are high impact
• Secrets in Vault, AWS Secrets Manager, or similar

---

Best practical stack for most teams

If you want a sane, production-ready default:

• Python
• FastAPI
• LangGraph
• OpenAI API
• Postgres + pgvector
• SQLAlchemy
• Jira/Linear/ServiceNow REST API
• Celery + Redis
• Pydantic
• OpenTelemetry/LangSmith

---

Simpler MVP stack

If you want to move fast with less infrastructure:

• Python
• FastAPI
• LlamaIndex
• Postgres + pgvector
• OpenAI API
• Direct ticketing API
• No queue initially, unless tasks are long-running

---

Key design recommendation

Make the agent tool-based, not fully autonomous:

• tool 1: search docs
• tool 2: query DB
• tool 3: draft ticket
• tool 4: create ticket only after validation/approval

That gives you much better reliability and control.

---

If you want, I can also propose:

1. a minimal MVP architecture,
2. a production architecture, or
3. a specific stack based on your ticketing system (Jira, ServiceNow, Linear, etc.).