Meta Tech ProviderInsights, deep dives,
product updates
Long reads on voice AI, flat-rate economics, Indian-language tooling and what we're building under the hood at Thinnest AI.
Tools, Not Skills: What Your Agent Can Do, and Why Every One of Them Starts Off
An agent that can only quote your documents is a smaller product than one that can look something up. ThinnestAI connects agents to 21 pre-registered MCP servers and a directory of 1,434 services — and ships every single tool switched off, with a hard cap of eight per agent. The cap is the feature.
There Is No Learning Loop: How an Agent Actually Gets Better Here
Nothing on this platform quietly rewrites your agent between conversations. An agent improves because you read what real customers said, then change one of four things: its knowledge, its instructions, which tools it may call, and when it hands over. That is slower than a self-improving system, and it is the only version of improvement you can audit.
How a Team Shares One AI Agent: Roles, Assignment, Escalation and Who Gets Told
There are no agent-to-agent execution modes here. One agent answers your customers on four channels, and the coordination that actually matters is between the people behind it — whose job a conversation is, who is allowed to change things, and who finds out when the agent hands one over.
Human-in-the-Loop, In Practice: How a Person Takes Over From the Agent
Most human-oversight features are an approval queue nobody reads. A business answering real customers needs something narrower: the agent recognises it is out of its depth, a teammate replies, and the agent goes quiet. That is what the shared inbox does, and this is exactly how it is enforced.
Four Ways In and Out: How ThinnestAI Talks to the Rest of Your Stack
Interoperability, without a new protocol to adopt. The agent calls MCP servers and your own HTTP endpoints. We push signed webhooks to you on five events. Your backend posts facts to the Events API. And your systems send order confirmations and one-time codes through a transactional messaging API. No agent card, no discovery network, and no agent-to-agent delegation — here is why.
“Calls Placed” Is Not a Metric: Measuring What an AI Voice Agent Actually Did
A campaign that dialled 5,000 numbers tells you nothing. Connect rate, average talk time, and the breakdown of why the rest failed — no answer, line busy, declined, not reachable — tell you what to change. Here is everything ThinnestAI records per call, and the one counting decision that keeps the connect rate honest.
Showing 19–24 of 28 posts