Research

This year’s focus will be on how organizations run when AI agents do much of the work. We call this AI-native data operations. What we learn will improve Business Protocol Management, the practice of running a business on a small hard core of protocols.

Premises

The research rests on three premises. We treat them as observations that could turn out partial or wrong, and the sessions and cases test them.

• Abundant cognition

  Reading and writing both get cheap. Processing unstructured data (PDFs, rate sheets, forms) and producing software cost a fraction of what they did.

  Tested in the water and construction cases

• Distributed agency

  Models act, and they run as many instances at once. Swarms of agents show up wherever work is done.

  Read against OpenAI’s incident report and Anthropic’s multi-agent study in theme I

• Mediation

  Information increasingly flows through models: people and agents find, read, and exchange it by way of LLMs and the protocols that connect them.

  Read against the Model Context Protocol in theme I and the Agent Payments Protocol in theme V

What we study

If those premises hold, running data operations means managing a swarm through its protocols. We study the three phases of Business Protocol Management. Each one relies on protocol vision, the capability to see the protocols a business actually runs on.

• See

  Seeing the protocols a business actually runs on, which usually go unnoticed until they break. Field logs record what happens and why, as it happens, so patterns show up. What an organization emits, and who can read it, shows where its protocols live. Nature helps too, because cells and flocks coordinate through signals their neighbours can read.

  Themes II, III and IV

• Design

  What agents can see and do, and the hard core that bounds them: permissions, data contracts, spending limits, and the checks every output must pass. That includes scoring agents against real outcomes and versioning their models, prompts and context.

  Themes I and V

• Evolve

  How discretion moves between people and agents, how it gets checked cheaply and recorded where others can see it, and how the hard core gets amended as the business grows.

  Theme VI

Open questions

These need research and partners. If you can help with one, get in touch.

1. Which few protocols should a company make hard, and how much freedom can it allow inside them? Companies rebuilding part of their business around AI, software architects, organization researchers.
2. How do we measure growth and prevented failures together? Reliability teams that track near-misses and delivery speed.
3. Where should decisions sit between people and agents, and how do we check them cheaply? Teams that run agents in production.
4. When should a free pattern become part of the hard core, and when should a hard rule soften? Platform teams, standards bodies, institutional designers.
5. What should a company publish, to whom, and how long should it stay current? Public agencies, data publishers, companies that share operational data.

Cases

Training

We train protocol vision outside the reading too. Members watch protocols at work, log them with the Protocol Bicorder and share them in #protocols-for-business, aiming at about a hundred over the year. The protocol watching guide explains how.

We also run workshops, listed under work so far, and we’re exploring a simulation in which a swarm of agents runs a small company and players set the guardrails. Anyone can help shape the reading plan by suggesting or challenging a reading.

Sponsor or partner

The Protocol Institute is an independent research organization, supported by the Ethereum Foundation through 2026 and becoming a Canada-based nonprofit. It is raising support for its 2027 programs, including this project. To sponsor the work or bring us a research problem from your domain, see how to support the Institute or get in touch.