Skip to Main Content

Prometheus on Gemini: a 128-agent swarm architecture for very deep analysis with Gemini 4. The task today's agents fail at: deep research you can actually trust. Single research agents stop early, cite pages that don't say what they claim, and never look for counter-evidence. You can't audit them. How it works: one question goes in. Gemini 4 Argon plans it into branches and sub-questions. Up to 128 Gemini Managed Agents (each its own Google sandbox with Search and page reading) work in parallel: researchers log every claim and every source, with the exact sentence copied from the page, into one shared evidence ledger. Researchers cross-check each other's key numbers from different publishers. Red-team agents re-open the sources themselves and file challenges; the owners must fix or drop the claim. A gate written in code refuses to let the team finish until every key claim has a confirmed red-team check, every challenge is resolved, every sub-question is answered, single-publisher numbers are labelled, and a random sample of quotes is really found on the cited pages (the system fetches them). Then Argon writes the report (per-branch section writers at 64/128 agents) and an independent Argon judge reviews it. Every claim in the final report carries a tag; click it to see its sources, the exact quotes, the red-team checks and the challenges. Live view shows the whole swarm working: searches, pages opened, sources, publishers, checks, challenges, tokens. Team sizes 8 / 16 / 32 / 64 / 128 agents: ~80% researchers in branches of ~12, ~20% red team, with a concurrency limiter and backoff so 128 agents run against the API safely. Evaluation (5 Sigma): the baseline is Google's own Deep Research agent (deep-research-max) answering the same question in the same run. We compare what can be verified, not how the prose reads: quotes re-found on the cited page, independent publishers per key claim, contradicting evidence surfaced, red-team checks, unresolved errors.

Team

More from Gemini 4 NYC Hackathon

We design an agentic system to automatically create + manage intelligent AI opponents/NPCs in video games. The suite of agents analyzes user gameplay…

Adaptive NPCs project preview

Manually routing PCBs is time-consuming and cost-inefficient. Current best algorithms (freerouting, PCBRouter) score very low on PCBWorld's hardest PCB…

routr project preview

Heliocam is a camera app where Gemini controls the camera. It watches what you're pointing at, sets the ISO, shutter, focus, white balance and zoom for you…

heliocam project preview

Repairs and assembly fail because manuals are lost, vague, missing, or nonexistent, so people break things or throw them out. KLIK turns a photo of any object…

Klik project preview

Iron Pulse turns your laptop into a 3D boxing ring. Fight an adaptive AI rival in real time using just your webcam. Throw real punches and lean to dodge…

Iron Pulse Arena project preview

Text your raw footage to an iMessage number and get a finished, trend-informed Reel back in the same thread.

Sticky project preview