What if AI agents had a Next.JS moment?
Every team building an AI agent today solves the same problems from scratch:
- How do I give it memory across conversations?
- How do I connect it to WhatsApp or SMS?
- How do I pass the user's identity into a tool call?
- How do I escalate to a human when it gets stuck?
- How do I test it when I change the prompt?
LangChain gives you primitives. You still wire everything else yourself.
There's no Next.JS for AI agents. No agreed structure, no conventions, no standard way to build. Every project looks completely different.
The idea is simple : you define what your agent can do, point it at a prompt, and pick a channel. Everything else is handled. Memory, PII redaction, human handoff, self-improvement.
I built a full AI receptionist running on WhatsApp to test it. 20 lines of code:
agent = Agent(
instructions="prompts/system.md",
skills=[check_availability, book_appointment, cancel_appointment],
memory=PostgresMemory(url=os.getenv("DATABASE_URL")),
learn_from_feedback=True,
privacy=True,
)
agent.run(channel="whatsapp")
Persistent memory, PII redaction, human handoff, and self-improvement out of the box. That's a full production AI receptionist on WhatsApp.
Still early. Would love feedback from anyone building in this space.
I think the “proof” point is the key one.
What we’re finding with OpsWatch is that once an agent is doing real customer-facing work, output quality alone isn’t enough. The useful evidence is whether the system can show the path from expected behaviour → observed action → exception/failure → human escalation → final outcome.
The interesting question is whether that proof should stay inside the product, or whether higher-risk deployments eventually need an independent layer verifying the evidence and behaviour separately from the team that built the agent.
Curious whether you’ve seen customers ask for that second layer yet, or whether they’re still mostly satisfied with internal logs/approval queues.