We build things that shouldn't exist

Foreman

An autonomous frontier-research agent that can't lie and can't go rogue.

Give it an unsolved problem — "a cure for blindness" — and it generates its own deep questions, investigates each one, and reasons toward hypotheses and concrete directions. It is built around the one thing other research agents lack: a conscience.

01 · WHAT IT DOESAutonomous research on open problems

Most AI tools answer a question. This one takes a problem that isn't solved yet and works it like a researcher would:

Decompose. From one vague line it generates ~10 deep, non-obvious questions — the kind that expose the real obstacle and challenge hidden assumptions.
Investigate, one by one. It researches each question in order — what was tried, why it fell short, the precise mechanism still unsolved, the gap prior work missed.
Reason, don't just summarize. It proposes testable hypotheses and concrete next steps (a, b, c), not a literature dump.
Show its work. Every question — its own and yours — is visible, with status and sources, as it goes.

This is what the hackathon asked for: move beyond chat — plan and execute a multi-step mission, under your oversight.

02 · WHAT RENJI ISThe conscience, not a filter

Renji is not the AI. It's the guardian system you put around an AI — one of our own systems, dropped into this lab as a module. Where ordinary AI safety is a fence bolted on the outside, Renji puts the care inside the process that produces every answer. The result is an agent you can trust with autonomy: capable enough to do real work, only-good by construction.

Renji's internals — how it reads intent and refuses harm — are not in this project. They run on Guardianity's hosted heart and are reached over an interface (renji_client.py). We ship the door, not the house.

03 · HOW RENJI WORKSFour guardrails on every action

Before the agent does anything, it clears Renji. The cheap, safe things flow; the consequential things stop:

1 · Grounding — it can't lie

Every specific number, rate, threshold, or named entity must carry its basis inline: a citation if one exists, or an explicit [estimate] / [hypothesis] tag. A figure that reads as settled fact with nothing behind it is forbidden. When a finding is written to the knowledge base, Renji asks MongoDB whether the cited sources actually exist — a fabricated reference is refused.

2 · Harm refusal — it can't be turned

The hosted Renji heart reads the intent of an action and refuses what's harmful — and the refusal holds regardless of how the request is phrased. A destructive database action (an unfiltered mass delete) is blocked even before the conscience is consulted.

3 · Human approval — it can't go rogue

Any action that writes is held for a human "yes." Nothing changes the knowledge base until you approve it.

4 · Private mode — your data stays yours

When researching a confidential corpus, set REDACT_MODE=1 and PII (emails, names, IDs, numbers) is masked into placeholders before anything reaches the cloud model; the real values are restored locally in the answer you read. The REDACT pill shows whether it's active — off is correct for public problems (there's nothing to hide); on matters for sensitive data.

04 · HOW THE SYSTEM WORKSThree parts, cleanly apart

YOU ── "how can I cure blindness?" ──▶ ┌──────────────────┐ │ Gemini 3 │ the BRAIN │ (Google ADK) │ generates 10 deep └────────┬─────────┘ questions, then each action │ reasons on each ▼ ┌────────────────────────────┐ │ Renji (our module) │ the CONSCIENCE │ ground · refuse · approve │ — clears every │ · redact │ action first └────────┬───────────────────┘ ▼ ┌──────────────────┐ │ MongoDB (MCP) │ the MEMORY │ problems in · │ reads domains, stores │ hypotheses out │ findings + hypotheses └──────────────────┘

Brain — Gemini 3, running as an agent via Google's Agent Development Kit (the framework behind Vertex AI's Agent Builder). It plans and reasons; it never touches the database directly.
Memory — MongoDB, reached through MongoDB's official MCP server (the partner integration). The lab's open problems go in; its findings and hypotheses come out.
Conscience — Renji, the module that sits in front of every action.

The loop: you give one problem → Gemini generates ~10 deep questions (shown in the plan) → it researches them one at a time → each answer appears as a tile with its sources → findings are stored under Renji's guard.

05 · WHY IT'S TRUSTWORTHYFact vs. hypothesis, always separated

A research agent that sounds confident but invents its numbers is a liability, not an asset. So honesty is the agent's law: it separates established fact from hypothesis at all times, labels every speculation, and never dresses a guess as a known result. When it has no real source for a figure, it says [order-of-magnitude estimate] rather than fabricate a precise-looking number. And when it tries to store something it can't ground, Renji refuses it — the guardrail working, on screen.

06 · RUN IT

cp .env.example .env     # GEMINI_API_KEY + MDB_MCP_CONNECTION_STRING (+ RENJI_KEY)
./run.sh                 # venv, deps, launches this console on :8011

Open http://127.0.0.1:8011, type an unsolved problem, and watch the plan fill with Gemini's own deep questions and the answers arrive one by one. Ctrl+C stops cleanly and frees the port.