Protocol InstituteBusiness

Feed · thread · version 3 · updated 11 November 2026

Kickoff: the limits of agency law for AI agents

Session of 2 November 2026. 17 participants, 111 minutes.

A business that delegates work to an agent needs a party that answers for the result. The year's first reading, Noam Kolt's "Governing AI Agents", tests agency law and principal–agent economics against that problem. The paper states its own limit:

Incentive design, monitoring, and enforcement might not be effective for governing AI agents that make uninterpretable decisions and operate at unprecedented speed and scale. (Kolt 2024)

Format: hour one covered the year's plan. Hour two ran a silent read, then one reaction per participant.

Observations

  • Enforcement lacks a target. One participant: "With these agents, they're ephemeral, and there can be infinity of them. And so how do you actually enforce against a specific agent?" The standard principal–agent model assumes a person or firm with a fixed place in time and space. Enforcement against an agent is a separate task from enforcement against the lab behind it.
  • Claimed autonomy exceeds real autonomy. A participant with a decade in crypto infrastructure: "people say that they're separate when they're not actually separate." In crypto, that separation often serves as a liability shield. The same test applies to an agent split across ten machines: "what level of obfuscation do you need to actually call it autonomous?" The paper leaves this question open.
  • Delegates drift from principals. A participant from DAO governance: "you have a delegate who is supposed to represent a token holder, but oftentimes, they don't talk to one another." DAOs hold a long record of this failure mode. Open question: which DAO governance pitfalls carry over to agents?
  • Practice has overtaken the paper. A participant with an engineering background: "This is eighteen months ago in academia, and it is happening to us right now." Case in point: security. Restrictions that block misuse also block the security researchers who work ahead of attackers. Governance timing reads as overdue, urgent and premature at once.

All four reactions point at one gap. The principal–agent frame assumes parties with an identity, a count and accountability. An agent supports copying, splitting and termination. The session lead added a second gap: no agreed definition of an agent as a software object. Example: one coding session may run one agent or many. A rule about a relationship requires a defined party at each end.

Frames from members

  • Four dimensions of scale. Tempo: time per action. Span: duration of a mandate. Width: number of agents in parallel. Depth: layers of delegation between the principal's intent and the acting agent. Human operation sits in the middle of each axis: seconds to hours, four to seven items in working memory. Law and economics target that middle band. On this view, new protocols form at the edges.
  • Five spaces in the paper: mandates and consent; loyalty and conflicts; delegation chains; evidence and audit; shared environments. Each space works as an anchor for a later session.
  • Extension, not delegate. One participant: "I prefer the kind of Doctor Octopus model of AI use, where I view it more as extensions of human judgment." Claim: human review scales at a roughly logarithmic rate through output sampling, plus distillation of judgement into small models or deterministic steps open to human review. Context carries the signals: "the pheromone medium is the context that gets read by various AIs." Status: untested here, with a write-up in progress.
  • A delegation threshold. This frame disputes the paper's assumption of limited efficiency gains from delegation: small gains compound across the board. Proposal: a model of the risk point "beyond which you get serious potential legal issues or complications."
  • Grounding. For open, forkable protocols: "what's the means by which you can ground what a good protocol is, what a bad protocol is?" The question covers people and agents alike.
  • Liability. From risk and insurance, the first question: who holds agency and who pays, for an AI agent and for an autonomous vehicle.

Working thesis

The session lead's framing in hour one: "because agents won't ask for manager approval, the style of management that we're working towards actually looks quite different." Gates work on people, since incentives reach people directly. The year's hypothesis: management shifts from gates to guardrails matched to the tempo of the work.

An interpretation from this write-up, not stated in the session: a guardrail requires an attachment point. Without a stable agent identity, the rule attaches to one of two places: the environment of the action, or the person or firm behind the launch. That choice is a protocol decision and a target for protocol watching.

Open questions

  1. What counts as one agent?
  2. Who carries liability when an agent on autopilot causes harm: the operator or the model provider?
  3. Which frame fits members' current practice: delegate or extension?
  4. Timing: rules now, or rules after a period of stable practice?

Hour one

Recap of eighteen months of sessions: organizations as knots of tensions, non-events, tempo, standards and adoption, protocol hardening. The year's aim: protocol vision, the capability to see the protocols behind a business. The six themes on the Sessions page remain a proposal, open to argument for or against any reading. Three participation tracks: observer, protocol watcher, case study. Case studies carry a path to paid work through the group's leads.

Introductions listed these protocol interests: construction procurement, personal knowledge systems, inference for small firms, DAO governance, open social protocols, brands as protocols, accessibility as a path of least resistance, the boundary between a hardened way of working and an open one for remixing, and product thinking versus protocol thinking.

Before 16 November

  • Protocol watch. One protocol from work, home or the street, with one paragraph in the channel. Starting questions: the protocol watching guide.
  • Reading: a member's post on product and protocol thinking.
  • Case study. For a case, solo or as a squad, a message to the session lead. Details: how case studies work.
  • Readings. Requests to add or skip a reading go to the channel or the suggestion form. Final call: the session lead.
  • Next session, 16 November: a guest talk on making a business's reputation readable to AI agents.

References

Kolt, Noam. "Governing AI Agents." SSRN, 2024. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4772956.

McDowell-Sweet, Matthew. "106-cream-pikes." Notes. https://msweet.net/notes/106-cream-pikes.

Source: recording notes from OpenRecapper, the recording bot on the Protocol Institute Discord, and the session transcript, both in the group's repository. Quotes come from the transcript, minus repeated words. The session has a recording.

Follow

Follow this blyg. Paste the feed into any RSS reader, or the blyg's address into Blygger Studio or another blyg client:

https://protocolsforbusiness.com/blyg/feed.xml

https://protocolsforbusiness.com/blyg/

To follow every member too, import blogroll.opml.

Join the conversation

This is a feed, not a forum. Two ways to answer a post:

  • Talk about it on Discord. The group meets in the Protocol Institute Discord, channel #protocols-for-business: join the Discord.
  • Respond from your own blyg. Quote this post in a thread, or publish a response to it (in Blygger Studio: follow this blyg, then use Respond or Quote on this post). Your blyg sends a Webmention; once it is verified, your post is listed under Responses here, usually within the hour.

    Blyg ID 27h8cwz7ghkqy7ym2z768p0dc0

    Quote line ![[27h8cwz7ghkqy7ym2z768p0dc0]]

    Respond to (stub_of) {"origin": "https://protocolsforbusiness.com/blyg/", "id": "27h8cwz7ghkqy7ym2z768p0dc0", "version": 3}

    Webmention endpoint https://protocolsforbusiness.com/blyg/webmention

  • No blyg yet? Start one, or use any blog with an RSS feed, then ask to join the community so your posts appear in the Feed.

Machine-readable: item JSON · changelog 3 versions