From the Lab

Voice Agent

one front door to all your agents

local-llmrouterstreaming
$ cat problem.md

The Problem

Once you've built a handful of agents and tools, the bottleneck stops being the agents and becomes the interface. Which terminal was that one in? Which command? Which bot do I text for which thing?

I wanted a single way in — talk (or type) like a normal person, and have the right thing happen — instead of remembering the wiring.

$ cat solution.md

The Approach

One conversational front door that sits in front of everything:

  • You say what you want in plain language
  • A small, fast model figures out the intent and routes it — to a tool, or to a more capable writing model
  • The response streams back as it's generated, so slow writers don't feel dead
  • Anything that posts or sends waits for an explicit confirm before it happens

The interesting part isn't any one model — it's the shape: a cheap router up front, capable workers behind it, and a confirm gate before any side effect.

$ cat impact.md

Why It Matters

That shape generalizes. Any team drowning in internal tools has the same problem — the capability exists, but reaching it is friction.

A routing front door with a confirm gate turns "ten tools nobody remembers how to use" into "ask for what you want." It's the difference between building agents and making them usable.

THE RECEIPTS

Cheap router, capable workers, confirm gate

you  →  small router model  →  pick the right path
                                ├─ tool call
                                └─ writing model (streamed)
                                    ↓
                             confirm before anything posts

Swap the models, keep the shape. The router stays small and cheap; the workers do the heavy lifting; nothing with a side effect fires without a yes.

← Back to the Lab