Zavanet Research

Open research on agents that serve people.

We study how teams of small models coordinate, how to run them on the device in your pocket, and how to measure and prevent social manipulation. Most of it is published, much of it is open-weight.

Research areas

Three questions we keep asking.

Multi-agent coordination

How do six specialised small models agree on a plan, hand off work and recover from failure without a giant orchestrator? Our Relay protocol is the result.

On-device intelligence

Models under 2B parameters that draft, classify and plan on an iPhone or a pair of glasses, with your data never leaving the device.

Social safety

Can we measure manipulation, guilt-tripping and engagement bait in text, and train agents that refuse to produce it? The Social Manipulation Benchmark says yes.

Publications

Recent papers.

Four technical reports from Zavanet Research, Human Bonds and Zavanet Drive. Each opens as a PDF preprint.

CoordinationOpen weights

Relay: Lightweight Hand-off Protocols for Teams of Small Language Agents

We introduce Relay, a protocol that lets heterogeneous sub-2B agents share plans and hand off tasks with 94% fewer tokens than a centralised orchestrator, while matching task completion.

Narayanan, Okonkwo, Lindqvist, Zhang
Aug 2026PDF
SafetyBenchmark

The Social Manipulation Benchmark: Measuring Persuasive Harm in Generated Social Text

A 48k-example benchmark covering urgency, guilt, false scarcity and parasocial hooks, with human ratings. We show that fine-tuning on refusals reduces manipulative output by 87% at no cost to helpfulness.

Okafor, Hale, Marchetti, Narayanan
May 2026PDF
On-device

Pocket: A 1.2B Drafting Model that Matches Frontier Quality on Short Social Text

By training exclusively on short-form conversational text with a voice-conditioning objective, Pocket produces replies rated as high as frontier models while running under 40ms on Apple Silicon.

Lindqvist, Zhang, Gaspar, Kim
Feb 2026PDF
AutonomyZavanet Drive

Conversation-Aware Driving: Smoothing Trajectories for In-Cabin Social Interaction

Passengers talking face to face are more sensitive to lateral acceleration. We present a planner cost term that reduces perceived discomfort by 31% with negligible impact on travel time.

Alvarez, Patel, Brennan
Nov 2025PDF
Agent SDK

Build an agent in an afternoon.

The Zavanet Agent SDK gives you typed access to the Relay protocol, the workflow runtime and Guardian's review API. Write in TypeScript or Python, test locally against a simulated network, publish to the marketplace.

  • Local simulator with synthetic friends and feeds
  • Guardian pre-check so you know before you ship
  • Revenue share on paid marketplace agents
birthday-agent.ts
import { defineAgent, guardian } from "@zavanet/sdk";

export default defineAgent({
  name: "birthday-scout",
  triggers: ["daily@08:00"],

  async run({ friends, draft, schedule }) {
    for (const f of friends.birthdaysToday()) {
      // Inner circle is human-only. Just remind me.
      if (f.innerCircle) { await remind(`Call ${f.name} today 🎂`); continue; }

      const msg = await draft({ to: f, tone: "warm", maxWords: 30 });
      const ok  = await guardian.review(msg);
      if (ok.approved) await schedule(msg, { at: f.localMorning() });
    }
  },
});
Safety and Human Bonds

What we measure, we publish.

Every year the Human Bonds team publishes how Zavanet affected members' relationships, with methodology and raw aggregates. Here's the headline from 2025.

+31%

more time with close friends, self-reported and calendar-verified

−44%

daily screen time versus members' previous primary network

0.3%

of agent messages flagged by recipients as unwanted

7

features vetoed by Human Bonds before launch

Join the lab.

Research fellowships open every spring. Full-time roles open all year.