Agents don't die
in the demo.
They die in production.

No — and the difference is the whole point. A boilerplate is code you copy once and own alone; the day you clone it, it starts drifting from everyone else's. agentFast is infrastructure: the durable-execution engine, the approval gate, the audit trail, the eval harness — the parts that are the same hard problem for every agent team, maintained in one place and improved for everyone at once. You still write your agent. This is the layer it runs on.

You should — agentFast runs on them, not instead of them. But an SDK is an engine, not a car. LangGraph gives you orchestration; it gives you nothing for the moment your agent restarts mid-task and loses the run, or fires a refund twice because no one approved it, or invents an answer no one catches. That gap between 'orchestrates' and 'survives production' is exactly what this fills — while you keep writing real LangGraph or real CrewAI underneath.

Production is the only thing it's for. It ships the parts that separate a demo from a deployment: durable execution (an agent that survives a process crash mid-task and resumes from the exact step), a human approval gate on risky actions that survives a restart, PII redaction in and out, cost and latency tracing on every run, a complete audit trail of every decision, and evals that fail your CI before a regression ships. It's built to sit in front of paying customers on day one.

Because the buyers who most need this — anyone touching money, health, or regulated data — can't send their prompts, tool outputs, and audit logs to someone else's cloud. agentFast runs entirely in your infrastructure. Your data never crosses your boundary, and you're not locked to a vendor's platform. That's a deliberate line against the hosted agent platforms: your intelligence, your infrastructure, your control.

It's source-available. You get the full source, own it, run it anywhere, and modify it however you like. Solo covers one developer for your own projects; Team covers up to 10 and adds client and commercial delivery. The one thing no tier allows is reselling agentFast itself — that's what funds the updates instead of the project getting cloned once and abandoned.

Any model — OpenAI, Anthropic, Gemini, Ollama, Groq — via a model-neutral layer, so you're never locked to one vendor and can self-host. Tools are MCP-native, in both directions: consume any MCP server the ecosystem publishes, and expose your own agent as an MCP server other agents can call. The 54 tools that ship are plug-and-play; adding your own is a one-liner.

Each SDK sits behind a thin adapter, so a breaking change hits a small, contained surface instead of your whole codebase — and fixing it once fixes it for everyone on that adapter. That's the quiet advantage of infrastructure over a copy you maintain alone: you're offloading the ecosystem-churn maintenance, not inheriting it.

It's being built in public right now, in stages, and tested hard as it lands — durable kill-and-resume is a real automated test against a real database with a real process kill, not a slide. You get the full source, yours to keep, plus 12 months of updates. I'd rather hand you something that survives production than rush you a half-tested skeleton.