agentFast vs LangGraph Starter Kit: Scaffolding vs Infrastructure
Credit where it's due: the LangGraph Starter Kit is a genuinely well-built project. Apache-2.0, free, clean modern dependencies, seven agent patterns, five LLM providers, a CLI scaffolder, and a typed Fastify server. If you want to stop rewriting LangGraph boilerplate in TypeScript, it does that job well — and its README says exactly that, without overclaiming.
This page isn't “we're better.” It's a different question: what happens when you kill the process mid-approval? That's where scaffolding ends and infrastructure begins, and it's worth being precise about, because both projects ship something called “human-in-the-loop.”
The same words, two different guarantees
The Starter Kit's HITL pattern calls LangGraph's native interrupt() directly inside the tool, and resumes with Command({ resume: decision }). That is the correct, idiomatic way to use LangGraph — and for a scaffolder it's the right call. What it doesn't give you is a guarantee about what happens on the resume path if the same approval is delivered twice, or if the box died between “approved” and “executed.”
agentFast treats that as the actual product. An approved action is claimed with a compare-and-set on executed_at in Postgres, so exactly one caller — across processes, across restarts, across a double-click on the approve button — can run the side effect. Refund once, not twice.
Side by side
| LangGraph Starter Kit | agentFast | |
|---|---|---|
| Category | Scaffolder / starter kit | Agentic infrastructure (trust layer) |
| License & price | Apache-2.0, free | Source-available commercial, from $89 |
| Language | TypeScript | Python |
| SDKs | LangGraph only | LangGraph, CrewAI, Claude Agent SDK, OpenAI Agents SDK, Vanilla |
| Patterns / templates | 7 patterns, CLI scaffolder | 3 working agents (Support, Research, Code) |
| Human-in-the-loop | Native interrupt() | Durable pause + exactly-once execution (CAS) |
| Survives a crash mid-approval | Not tested | ✓ Automated test: real Postgres, real SIGKILL, two processes |
| Default persistence | In-memory; Postgres optional | Postgres by default |
| Long-term memory | In-memory store (even with Postgres set) | Persistent pgvector recall |
| Audit trail | — | ✓ Who approved what, when, and what followed |
| API auth | None | ✓ Bearer token on every endpoint |
| Guardrails / PII redaction | — | ✓ Redaction, safety filters, rate limits |
| Evals in CI | — | ✓ Thresholds fail the build |
| MCP | Consumer (loads external MCP tools) | Two-way — consume, and publish your agent as an MCP server |
Three things worth knowing before you ship
These aren't criticisms of a free scaffolder — they're the specific gaps you'll need to close yourself if you take it to production.
- Memory doesn't survive a restart. The cross-thread store is in-memory even when Postgres is configured. Checkpoints persist; long-term memory doesn't.
- The HTTP server has no auth. Routes for invoke, stream, and resume bind to all interfaces with no token, no rate limit. Anyone who can reach the port can approve a pending deletion.
- No guardrails, no evals. There's no PII redaction before the model sees input, and nothing to fail your build when answer quality regresses.
Which should you use?
If you're in TypeScript, exploring LangGraph patterns, and want a fast, free, clean start — use the Starter Kit. It's good at that, and nothing here suggests otherwise.
If your agent is about to take an action that costs someone money — a refund, a payout, a PR merged into main — the question stops being “which scaffolder” and becomes “what happens when this breaks mid-flight.” That's the layer agentFast exists to be: durable execution, an audit trail, and evals in CI, self-hosted in your own infrastructure, on whichever SDK you already chose.
Comparison based on a read of the public repository (Apache-2.0) in July 2026 — README, package.json, src/apps/interrupt.ts, src/config/checkpointer.ts, src/tools/mcp.ts, and src/server/index.ts. Open-source projects move fast; check the repo for the current state.
agentFast is the production layer — memory, 50 tools, observability, evals, guardrails, human-in-the-loop — for LangGraph, CrewAI, the Claude & OpenAI Agent SDKs & Vanilla. Own it for life.
Get agentFast — from $89