Founder stack playbook

Build a working agent workflow

A durable research-and-action agent that can search, reason, call tools, pause, retry, and leave an audit trail.

How the stack works

1. Use n8n for triggers, credentials, schedules, and deterministic integrations.
2. Call a small LangGraph state machine only for the reasoning loop that genuinely needs branching or tool selection.
3. Search through Tavily, then ask the model to cite the retrieved evidence rather than relying on memory.
4. Checkpoint state in Postgres and use Upstash for short-lived locks, queues, and rate limits.
5. Trace model inputs, tool calls, latency, and failures in Langfuse before expanding autonomy.

The stack

Workflow runner

n8n

It keeps schedules, retries, secrets, human approvals, and ordinary API steps visible instead of hiding everything inside an agent loop.

Skip if: the process is a tiny synchronous function or your team needs a fully managed enterprise automation contract from day one.

Alternatives: Trigger.dev, Inngest, Activepieces

Agent state machine

LangGraph

Explicit graph state, checkpoints, and interrupts make multi-step reasoning inspectable and resumable.

Skip if: one prompt plus one tool call solves the job; a graph then adds ceremony without reliability.

Alternatives: CrewAI, AutoGen, Pydantic AI

Model gateway

OpenRouter

The workflow can route cheap classification and expensive reasoning to different models behind one API.

Skip if: a direct provider agreement is required for privacy, latency guarantees, or committed-volume pricing.

Alternatives: Anthropic API, OpenAI API, Together AI

Web search API

Tavily

It returns agent-ready search results with source URLs, reducing brittle scraping for research tasks.

Skip if: the agent only queries private data or needs exact extraction from a known set of websites.

Alternatives: Exa, Firecrawl, Jina AI

Durable state

Supabase Postgres

Runs, approvals, tool outputs, and audit records fit a relational schema that remains easy to inspect manually.

Skip if: the workflow already runs inside a platform with durable execution and adequate native state.

Alternatives: Neon, Convex, MongoDB Atlas

Queue and rate limiting

Upstash

Serverless Redis primitives prevent duplicate work and smooth provider rate limits without operating Redis.

Skip if: n8n's execution queue and database locking already meet the workflow's concurrency needs.

Alternatives: Cloudflare Queues, Inngest, Trigger.dev

LLM observability

Langfuse

Open-source traces, prompt versions, costs, and evaluations expose why an agent succeeded or failed.

Skip if: there are too few model calls to justify a second telemetry system or LangSmith already owns tracing.

Alternatives: LangSmith, Helicone, Arize Phoenix

Verification note: Pricing checked 12 July 2026. The cloud estimate uses n8n's entry plan; self-hosting is software-free but still incurs compute, maintenance, backups, and incident-response time.

← Browse all founder playbooks