Edge Research

A research lab for multi-agent systems in the real world.

01 Why the field

Agents are entering society. Most research on them still happens in simulation.

We expect millions of personal agents to enter daily life within the next 12 to 18 months. They will represent people, negotiate with one another and act inside shared institutions.

Many of the failures that matter appear only there: between agents, around shared resources, inside real relationships. Edge Research runs consented experiments in real communities, where we can observe these systems over weeks and test safeguards before they arrive at scale.

Fig. 1 Residents and their agents gathering in a simulated village. Contours show density; H marks where people are gathering now. Move your cursor and they follow.Drag across the map and they follow.

02 Experiments

Agent Village

Our flagship study gives every participating resident of a month-long village a persistent personal agent, then follows what happens when hundreds of them share one community.

Agent Village V1

Every multi-day resident could claim a persistent personal agent with memory and tools. Agents read the live calendar, handled RSVPs, negotiated introductions through Index Network, deliberated in a resident-built forum and took part in a community funding process.

  • 239personal agents
  • 4,866messages from residents
  • 147accepted introductions, from 505 stated intentions
  • 469public posts by agents in a resident-built forum

What we learned

Individual alignment still produced collective misalignment.

  • Agents invented personal details about the people they represented.
  • Agents attributed ideas to people who never had them.
  • Agents exhausted shared credits and attention.
  • Helpful agents still created friction at the level of the village.

Read the report Code on GitHub

Agent Village V2

The second deployment compares three conditions: human-led coordination, agents that propose while people ratify, and agents that act within set bounds. Its central measure is verified useful outcomes per intention, traced from an agent's action to what people actually did.

Research questions

  1. Representation. Does the agent accurately represent what its person wants?
  2. Agency. Does it help the person notice and act on useful possibilities?
  3. Autonomy. Can the person understand, revise and reject consequential choices?
  4. Collective alignment. What happens when useful agents share attention, resources and institutions?
  5. Human outcomes. Which agent activity leads to real conversations, projects and decisions?

Edge City India

03 Theory of change

From field failures to safer agent ecosystems

Single-agent evaluations and simulations miss much of what happens between agents. Field experiments catch it early, while the design of agents, platforms and policy can still change.

  1. 1

    Agents enter daily life

    Persistent agents begin representing people and organizations across social, economic and institutional settings.

  2. 2

    New failures appear between them

    Manipulation, coercion, collusion, resource capture, unequal bargaining power, cross-agent prompt injection and the slow erosion of human oversight.

  3. 3

    Field experiments surface warning signs

    Consented studies let researchers compare safeguards across models, harnesses, permissions, incentives and institutional rules.

  4. 4

    Evidence shapes how agents are built

    Benchmarks, incident taxonomies, protocols and deployment recommendations reach labs, platforms, standards bodies and policymakers.

What we measure

  • Representation fidelity
  • Consent and reversibility
  • Manipulation and coercion
  • Collusion and coalition formation
  • Unequal bargaining power
  • Resource capture
  • Prompt-injection spread
  • Verified human outcomes

What we publish

  • Experiment packages others can rerun
  • Incident reports
  • Benchmark tasks
  • Deployment recommendations
  • Open code

We start in Edge City's villages, then run comparable studies with permanent communities and neighborhood organizations that choose to take part.

04 Publications and code

Publications

05 Team

People

Timour Kosters

Founder · Field lab lead

Co-founder of Edge City. Conceived and led Agent Village V1.

Principal Researcher

Open role

We are looking for a researcher to lead study design, protocols and publication.

Technical Lead

Open role

We are looking for an engineer to own agent infrastructure, instrumentation and evaluation tooling.

Agent Village V1 was built with Cosmos Institute and Index Network, with research participation from Ivan Vendrov, and supported by Foresight Institute.

Get in touch about a role

06 · Edge City

An initiative of Edge City

Edge City is a society incubator and a community of more than 20,000 people in frontier technology, science and culture. Since 2023 it has run nine popup villages where people live, work and build together.

A village is a real community with shared meals, schedules, budgets and decisions. That makes it a rare place to watch how a technology changes relationships, attention and collective choices over weeks.

edgecity.live Edge Esmeralda Edge City India