Meet Ron

Notes and research questions from Ron Kagan

Personal agents: what changes when we delegate?

  • Can our context travel with us when we change agents?
  • What happens as specialized knowledge becomes easier to access?
  • Can a business work effectively with a customer’s agent?
Go to the resourcesGet email updates

Keep the context. Choose the agent.

The more an agent knows about our work, preferences, and decisions, the more useful it can become. That also raises the cost of switching providers.

I’m interested in context we can inspect, edit, and carry between tools. Plain-text notes and Markdown offer a practical starting point. The harder questions are what context to retain, who can access it, and how to avoid carrying stale assumptions forward.

Provider choice matters more when changing tools also means rebuilding the history that made them useful.

I’ve written about making my own AI agent replaceable, including what transfers and what still needs work.

Domain knowledge is being compressed.

A buyer can ask more sophisticated questions about pricing, integrations, security, or procurement with an agent’s help. The effort required to bring specialized knowledge into a decision is changing.

The question I want to examine is where that access becomes reliable judgment. Can an agent recognize which knowledge applies to this company, these constraints, and this decision? Can it identify what it doesn’t know?

Research discussed by Harvard Business School gives us a useful starting point. AI assistance has improved performance on some professional tasks and worsened it on others. That makes task-level testing especially important.

The customer’s agent meets the whole business.

A buying decision can touch marketing, product, pricing, security, procurement, onboarding, and support. An agent needs accurate information and a workable next step across those handoffs.

Helping me research a choice is different from acting for me. Before an agent sends a message, shares data, or makes a purchase, the scope and approval need to be clear.

Poppy, the draft Personal Agent Protocol, is worth following here. It proposes a way for personal agents to identify themselves and interact with businesses through websites, APIs, and company agents, within the customer’s permissions.

Its focus is agent–business interaction. Portable personal memory remains a separate question.

A test worth running

How far can an agent get through a realistic B2B evaluation before it needs a domain expert?

One study I’m considering would give agents realistic buying briefs, compare their answers with a human-expert baseline, and track accuracy, missing evidence, unsupported conclusions, and useful next steps. Then change how a controlled vendor site presents the same facts and repeat the test across models.

This is a proposed experiment. Its value would be showing which changes improve evaluation and where expert judgment still matters.

Resources I’m reading

The proposals, specifications, and research behind these questions.

Sierra · Introduction

Introducing the Poppy draft

An accessible explanation of agent–business interaction and the protocol’s design goals. A proposal from its creators, rather than evidence of widespread adoption.

Personal Agent Protocol · Draft 0.1

Read the specification

The technical source for discovery, sessions, permissions, and interaction channels. Useful for understanding what Poppy actually specifies.

Personal Agent Protocol · Optional draft

Actions, approvals, and retries

The operations extension shows how proposed actions, approval, tracking, and retries fit together. A draft design to evaluate in real implementations.

ACES · Research paper

What Is Your AI Agent Buying?

University-affiliated experiments on shopping agents’ choices and biases in controlled consumer marketplaces. Useful evidence about agent behavior; it doesn’t establish B2B adoption.

Harvard Business School · AI at Work

Where AI helps, and where it falls short

Research showing task-dependent gains and failures in AI-assisted consulting work. A reason to test judgment carefully, without assuming expertise transfers uniformly.