Finalist
Firn
Founders want to grow their business with AI, but they can't delegate to agents they can't trust.
Core trick
Extracts atomic facts from documents and calls, keeping only those whose quote matches the source verbatim per SQL check, then maps conflicts per zone.
Description
Founders want to grow their business with AI, but they can't delegate to agents they can't trust. The company's knowledge is scattered across CRM, chats and call recordings; sources contradict each other and much is simply missing. Agents either hallucinate confidently or stall waiting for clean data. Our pilot founder (RecPlace, a geomarketing company for SMBs) spent two months building his own AI "second brain" on top of all his systems — and ended up babysitting it, re-checking every answer it gave his team.
Firn turns that chaos into a provable map of the business and compiles digital employees from it, all inside Snowflake:
Distill — Claude (via Cortex AI_COMPLETE) reads every document, chat and call and extracts atomic facts; a fact counts only if its quote is found verbatim in the source, checked by SQL.
Map — facts land in 9 business zones / 37 subzones; Firn computes, per zone, what is known, what conflicts and what is blind, and finds contradictions on its own (fact vs fact, fact vs CRM, rule vs practice).
Compile — COMPILE_AGENT turns a founder-written role plus the map into a Cortex Agent: mandate per zone, governing norms with authority, goals and metrics.
Act under uncertainty — the agent (Firn Chief of Staff) answers with evidence and the edge of its knowledge, and proposes actions tagged ACT / HYPOTHESIS / ASK / APPROVE, with money at stake, a loss limit and a review date. Nothing is executed without approval; every step is traced.
tvpustovoy/reliable-snowflake-agents-factory★ 00 forksHTMLCreated Oct 23, 2026 · last push Oct 23, 2026
README excerpt
# Firn — Snowflake Agents Factory
[](https://youtu.be/FquE-ECczJM)
▶ **[Watch the 60-second demo](https://youtu.be/FquE-ECczJM)**
> **Know what you know. Act anyway.**
> Firn grows business agents inside Snowflake from a company's own messy data —
> agents that show their evidence, map what they don't know, and still act safely under conflicting, incomplete knowledge.
*Anthropic × Snowflake Hackathon, Oct 2026.* · **Live build status: [docs/STATUS.md](docs/STATUS.md) · [HTML](docs/STATUS.html)**
## The problem
Founders want to grow their business with AI, but their knowledge is scattered across CRM, chats and calls. Agents can't be delegated to, because real knowledge is conflicting and incomplete: agents either make things up or wait for clean data that never comes.
**Teams that can trust their agents delegate more. Teams that delegate more capture the most value.**
## How it works
```
feed everything → distill a provable map → compile agents from the map → agents act → outcomes → better map
(RAW/CORE) (facts + quotes, (mandate per domain: (ACT / ASK / (rewards,
conflicts, gaps) ACT · DRAFT · ASK) DRAFT→APPROVE) evolve vN+1)
```
- **Map of knowledge** — every fact is a verbatim quote verified by SQL; conflicts and gaps computed per business domain.
- **Agent factory** — `COMPILE_AGENT(mission)` turns the map into a Corte