Work/Agar
A Minecraft companion that talks at human speed and acts while it's still talking.
Overview
Agar, from Agartha Labs, is a Minecraft companion you talk to like a duo partner. It answers in about the time a person would, starts acting while it is still talking, and remembers you between sessions. We built it as team larper at AGI Summit 2026 in San Francisco, and it took the Insforge prize.
I worked on the voice layer and the async tool calls: the part that lets it say "aight, otw" and already be walking.
- When
- Jul 2026
- Where
- AGI Summit 2026 · Insforge prize
- Role
- Voice and async tool calls, team larper
origin.
Two working halves that couldn't reach each other.
The pieces already existed. The version before Agar took 6.5 to 13 seconds to say a word, its voice layer could talk but had no tools loaded, and its action layer pushed speech into a queue nothing drained.
Agar is the rewrite that takes everything slow out of the path to speech.
voice.
One live speech session, and you can talk over it.
The conversation lane holds a persistent Gemini Live speech-to-speech session with native voice activity detection and barge-in, so you can interrupt it mid-sentence like a person.
Its function calls dispatch straight to the Mineflayer body over a local MCP connection in single-digit milliseconds.
async tools.
Say it, and it is already moving.
A goal given by voice returns immediately and reports completion later, so "aight, otw" lands while the bot is already pathing, and "got the wood" arrives when it's actually done.
On demo day the first tool call (1763 ms p50) beat the first audio (1828 ms p50). That ordering is what makes it feel like a partner instead of a chatbot with a skin.
system.
Three lanes, and nothing slow sits in front of speech.
A reflex lane runs at 15 Hz on a 66 ms budget with no language model at all: follow, safety, lava, auto-eat. A reflection lane does memory retrieval, fact writeback and deeper reasoning asynchronously, never blocking a word.