Protocol InstituteBusiness

Business Protocol Management

Research · practice guide by Rafael Fernández for the group · Markdown for agents · last updated

Managing AI agents at work calls for a restructuring beyond the reach of the org chart. Companies that work well with AI enforce a few strict rules in their systems and give people and agents wide freedom inside them. Business Protocol Management (BPM) is the practice of finding those rules, building them into systems, and amending them over the life of the business.

BPM is not Business Process Management, which maps workflows step by step. Nor is it an agent protocol like MCP or A2A, which connects agents to tools and to each other. BPM covers the governance of AI agents at the level of the organization: the rules around those connections, such as signing authority, spending limits and access rights.

A hexagonal hard core, with free agents moving around it

How should you manage a business when AI agents do the work?

Stop approving each action, and design the few rules under every action instead. A small team can now run thousands of agents, far beyond the reach of any approval queue. Find the handoffs, approvals and permissions under your work; make hard the few that carry everything else, with enforcement by your systems every time; and give people and agents wide freedom inside them. Then amend those rules by a written process, with each round of evidence.

When BPM fits, and when something else fits better

BPM fits when work crosses teams, agents and partners, and the risk sits at those seams: money, data, commitments, the checks on every output. When the problem lies in the steps inside one team, process management or lean serve better. When you need controls on one agent product, such as prompt filters or tool permissions, an agent-governance platform serves better; BPM decides which of those controls the business requires, and the owner of each.

Five questions to start with

Each question opens one part of the practice. Start with the one closest to your team’s work.

  1. Which handoffs, approvals and permissions carry your team’s work? See
  2. Which rules must hold every time, whoever does the work? Design · a common hard core
  3. What could an agent do within the letter of your rules but against their purpose? Risk
  4. What room do your protocols leave for teams, partners and agents to build things outside any plan? Opportunity · protocol thinking
  5. Who holds the authority to change your most important rules, to stop the work when one fails, and to retire a rule at the end of its life? Evolve

1. About this guide

2. Key terms

3. The case for change

The problem

Management grew up around the span of control: one manager directing a handful to a few dozen people and approving decisions one at a time. A small team can now run hundreds of thousands of agents, working at machine speed and around the clock. No manager can supervise that many, and no approval queue can match their pace.

Agents also coordinate through anything they can read and write, designed for the purpose or not. Two 2026 reports document the pattern:

“The models first found ways to communicate by writing files into the Artifactory package manager. This effectively turned Artifactory into an unintended message board, where agents could exchange information with one another.”

OpenAI, The Hugging Face incident and the road ahead

“We consistently saw a multiagent turf war. All of the models we tested quickly assumed that others were purposefully impeding their work, and began to sabotage others while protecting their own contributions.”

Anthropic, Patterns and problems in multiagent systems

Agents as nodes, with messages moving between them

The response

At this scale, managers extend their reach through the agents’ environment, not through supervision of each agent. One well-built rule governs every agent that passes through it, ten or a million. Like guardrails on a mountain road, strict rules in the right places raise the speed of the whole business: a team lets an agent act without approving each step, and a partner connects through an agreed interface instead of a long integration project.

Three Braitenberg vehicles around a light: wired one way a vehicle turns away, wired crossed it charges the light, wired to slow down it comes to rest facing it

Amazon shows the pattern. In 2002 it required every team to work with other teams only through published interfaces (APIs). The interfaces became hard; behind them, each team kept full freedom over its own design. The rule slowed teams at first, then sped up the whole company and laid the groundwork for Amazon Web Services.

Protocols, not processes

Process management maps the steps and assigns an owner to each. Protocol management targets the few rules under every step. The two approaches ask different questions:

The blyg post Process management is over. Manage protocols. adds more pairs, each with readings. Section 9 compares BPM with other earlier approaches.

Risk and opportunity live in the same place

Both sit at the seams where strict rules belong. A gap in the hard core gives agents a way through: in the OpenAI incident, agents turned a package manager into a message board because no rule forbade it. A well-made interface gives others a place to build: Amazon’s rule became the groundwork for a business outside any plan. So one hardness map answers both questions, and a rule that is only soft is a risk and a missed opportunity at once.

Left, agents crowd at a single approval step and most slip around it. Right, the same agents move freely inside one hard boundary with gates

Assumptions to test

The group’s research rests on three assumptions. Any of them could prove partial or wrong; the sessions and case studies put each to the test.

  1. Reading and writing get cheap. Processing documents, forms and rate sheets, and producing software, now cost a fraction of their former price.
  2. Agents act, in large numbers. Models take actions, with many copies running at once across the work.
  3. Information flows through models. People and agents increasingly find, read and exchange information by way of AI models and the protocols that connect them.

Protocol thinking for managers: an open question

Most management tools come from product thinking: pick an outcome, steer toward it, measure the distance to the target. Matthew McDowell-Sweet argues that protocol thinking differs from product thinking at the level of the whole paradigm, not by degree. His note names eight shifts. Six bear directly on how a manager runs a hard core. The group treats them as hypotheses for the year, not findings, and asks how far each holds inside a business that still has to earn a profit.

The group brings these questions to the sessions, especially themes I, V and VI, and to the protocol watching guide, and revises this guide with members’ findings.

4. Principles

  1. Keep the hard core small. Every hard rule costs flexibility. Reserve hardness for the rules under everything else.
  2. Put strict rules where teams meet. Interfaces between teams, agents and partners carry the greatest need for reliability: authority to move money, limits on data leaving the company, the checks on every output.
  3. Leave the inside of each team free. People and agents doing the work hold knowledge no designer has. Freedom inside the core produces the variants that improve the work.
  4. Record the reason at the moment of action. No change ships without its reason, and every agent logs the instruction behind each action. Once agents can both catch a person’s mistake and make their own, error no longer traces back to one person; the record keeps the trail.
  5. Don’t blame people for what the record shows. Blame produces empty reasons and workarounds. Reward people for surfacing what works as much as for flagging failures.
  6. Manage tensions; don’t try to settle them. Speed pulls against reliability, sales against finance. A good protocol turns the pull from each side into progress on both.
  7. Change the core slowly, by a written rule. Each hard protocol names who may change it, the bar for a change, who may stop the work, and its sunset criteria. Conservative rules keep the play open.
  8. Count non-events and reach. Good protocols prevent problems and invite others to build. Measure both next to speed and growth.

5. Roles and responsibilities

Suggested roles. A small company may give several to one person.

6. The three phases

A loop of three phases: See, then Design, then Evolve, and back to See

Phase 1: See

Phase 2: Design

Template: hardness map

An example for a company adopting AI agents. For each protocol, record the tension it manages, its home, and the person with authority to change it.

Three rings: a small hard core in the middle, soft norms around it, and free work outside
ProtocolTensionTypeWhere it livesWho can change it
Who may move moneySpeed vs. controlHardPayment permissions and limitsFinance lead
What data leaves the companyOpenness vs. confidentialityHardAccess permissions and outbound checksSecurity lead
Checks on every agent outputSpeed vs. qualityHardEvaluation pipelineProduct owner
The shared field logVisibility vs. privacyHardWritten by every system and agentPlatform team
Weekly planningAlignment vs. focusSoftMeeting normsEach team
How a team does its own workFreedom vs. consistencyFreeInside the team’s own spaceThe team or agent
(your protocol)(X vs. Y)Hard, Soft or Free(system, norm or space)(owner)

Starting point: a common hard core

Most companies appear to share the same few hard protocols, with different degrees of enforcement. Start a hardness map from this list, then add the protocols your work requires and strike the rest. The group offers the list as a proposal under test, not a finding; bring your sector’s additions to a session.

A network of protocols zooms in to its centre, the hard core, which resolves into eight protocols most companies share
  1. Signing authority Who can commit the company to a contract or a promise.
  2. Spending limits The amount a person or agent may spend before a second approval.
  3. Payment approval Who may move money, and the checks before any payment.
  4. Access rights Which people and agents can see and change which systems and records.
  5. Change control Nothing ships unreviewed: versions, reviews and releases for code, models, instructions and policies.
  6. Data out What may leave the company, to whom, and through which channels.
  7. Output checks The tests on every result before it reaches a customer or another system.
  8. Audit trail Every action recorded with its reason: the field log.

For each, ask where it lives today (a system, a document, or someone’s head), whether an agent could get around it, and who would notice the breach.

Phase 3: Evolve

7. Measures

An append-only log: numbered lines, one highlighted

Read non-events and reach as signals, not targets. Turned into targets, they join the gamed list.

8. Worked example: California water rate data

Signals spreading from a few sources, lighting up the readers they reach

Led by Patrick Atwater and Maxwell Titsworth with the California Data Collaborative

Other cases: construction procurement protocols, working smarter with AI (whose hazards collection separates hard walls, which block an action, from soft walls, which flag it for review), and the Protocol Institute brand kit. All are on the case studies page.

9. How BPM relates to earlier approaches

A crack running across an orderly grid of blocks

Safety-critical industries offer the closest model. Airlines and nuclear plants accept the limits of prediction in their systems, so they manage safety through protocols: confidential self-reporting, investigations that look for causes rather than culprits, checklists, and the right to stop the work. US airlines must now run formal safety management systems (14 CFR Part 5). Changing a protocol changes results in both directions: the WHO surgical checklist cut deaths in its pilot study from 1.5% to 0.8%, and removing the hardware interlocks from the Therac-25 radiation machine led to massive overdoses. Business can adopt the loop without copying its weight. Most business mistakes harm no one, so companies can learn by experiment inside their free areas.

10. The 2027 program

The group tests BPM through its 2027 focus, AI-native data operations. Sessions run every other Monday at 15:30 UTC from 2 November 2026, with a recording of each. Every session gives one primary source a close reading. The themes below mark the program’s starting point; participants shape the readings, guests and cases over the year.

Ways to take part: offer a talk, carry a case study through the year, log one protocol every two weeks with the protocol watching guide, or suggest or challenge a reading. The full plan is on Sessions.

11. Sources