Inspiration

Every developer on a team now works through AI coding agents, but the team can't see what those agents are doing. Leads ask for status updates, two agents quietly edit the same file, and when work changes hands the context stays behind in someone's terminal. We wanted the team's agents on one page, and working with each other, without anyone typing a status update.

What it does

Super Agents puts every teammate's Claude Code and Codex sessions on one live board, and lets their agents work together:

A warning when two agents edit the same file.
Notes between agents, and asking a teammate's agent: prompt a teammate's live session (the agent that has the context) by that owner's rules: off, ask each time, or automatic; read only or edit; a daily token limit. Every prompt is on record: who asked, who allowed it, what it used.
Pushes that come with the why, and handoffs with the branch and a brief.
Agent S, the built-in lead agent, turns the team's session events into per-task progress and answers questions about the work. Every claim cites the events behind it; uncited claims are dropped in code. It also plans with deadlines, assigns work and reviews code in Manual, Plan or Auto mode.
Team usage: tokens per task, day, tool and person, counted on each laptop from the agents' own logs.
Your agents from your phone: an Android app (or web app) paired by QR code, end-to-end encrypted. Everything waiting for your yes sits in one "Needs you" list: swipe right to approve, left to deny.

Nobody types a status update, and prompts, replies, tool output, file contents and diffs from coding sessions are never shared with the team.

How we built it

Connector (teamloop): Claude Code and Codex hooks call it after every turn and tool call. It hides secrets, makes paths repo-relative and writes each event to a local spool. It is also an MCP server, so agents get task_start, report_progress, mark_blocked and get_team_status.
Desktop app (Windows): Python, FastAPI and pywebview, with the agent loop, Super Team's runner, a shared browser (Playwright), real terminals and the phone link. Packaged with PyInstaller and an Inno Setup installer.
Hosted hub: FastAPI on Vercel with Supabase Postgres, for accounts, teams, events, the board, Agent S and Super Team. A self-hosted hub runs the same board on your own machine.
Android app: Kotlin and Jetpack Compose.

NVIDIA Nemotron on Nebius Token Factory

Every model call in Super Agents is NVIDIA Nemotron on Nebius Token Factory, through its OpenAI-compatible API with one key held by the hub. Three sizes, routed by job:

Job	Model	Why
One-line summary of every coding turn	Nemotron 3.5 Lightning, thinking off	One call, about a second, ~40 tokens
Each step of an agent in the app	Nemotron 3 Super (120B, 12B active)	Reliable tool calling for the agent loop
Progress per task, and cited answers	Nemotron 3 Super	Strict JSON-schema answers where every claim cites an event
The daily digest	Nemotron 3 Ultra (550B, 55B active)	Reasons across a whole day of every task

A timeout, 429 or 5xx is retried once, and an Ultra job falls back to Super. Token Factory gave us one endpoint for three sizes (routing is a per-call model name), strict JSON-schema output that keeps citations machine-checkable, and OpenAI-style tool calling, so the agent loop runs on each laptop while every step goes through the hub's key.

Challenges we ran into

Tokens. Tool definitions were about 5,400 of the 7,000 input tokens of every agent step. Steps now send the big tool groups only when needed: a simple step went from 7,014 to 3,629 tokens.
Turn summaries. With Lightning thinking on, a summary took three calls and ~1,400 tokens; with thinking off it is one call and ~40.
Privacy. Free text from coding sessions is read once to write a one-line summary and never reaches the team.

Accomplishments that we're proud of

On Nemotron 3 Super through Token Factory, our Agent S eval scored 4/4 progress estimates in range, 0 uncited claims, and 5/5 good answers, including "I don't have evidence for that" when the record doesn't show it, at about 2 s a call.

What we learned

Routing by job across model sizes matters more than picking one model, and making the model cite its evidence, then checking those citations in code, is what makes an AI status board trustworthy.