AI Virtual Inn of Court
Introduction to Agents
Ethan B Holland - April 2026 - https://ethanbholland.com/
Agents v. Smart Chatbots
Even without agents… we’ve got powerful sidekicks
The ChatGPT home sale.
In late 2025, Robert Levine, CEO of consulting firm ComOps in Cooper City, Florida, decided to sell his home without a realtor. During a holiday road trip, he and his wife began prompting ChatGPT about the entire home-selling process. The AI handled pricing strategy, marketing plans, room-by-room renovation recommendations for ROI, scheduling of showings, and negotiation guidance. Real estate agents had estimated his home's value at roughly $854,800. ChatGPT encouraged him to list $100,000 higher. The home sold for $954,800 — one of the highest per-square-foot prices in the area — and closed in five days. Levine showed the property to 15 prospective buyers personally, one-third of whom submitted offers. He still hired a real estate lawyer for the legal work. The story ran in Fortune on March 21, 2026.
Agents v. Smart Chatbots
Even without agents… we’ve got powerful sidekicks
The AI tax preparer… closer to an agent (but mostly a saved system)��On April 2, 2026, Perplexity AI launched "Computer for Taxes" inside its Perplexity Computer agent platform ($17/month Pro subscription). Users upload W-2s and 1099s, and the AI drafts federal returns on official IRS Form 1040 with supporting schedules. Critical nuance: it does not file taxes — it drafts and reviews them. But in Perplexity's own testing, the AI caught a 67% understatement of deductions on an attorney-prepared return under the 2025 "No Tax on Overtime" provisions. A PYMNTS survey found 25% of U.S. workers planned to use AI for tax help in 2026, more than double the prior year. Among Gen Z, 62% are open to AI financial planning.
What is an AI agent, really?
Scripted Pipelines Versus Autonomous Agents
A chatbot is reactive: it waits for your question and gives you an answer from its training data.
An AI agent is proactive: it can perceive a goal, plan steps to achieve it, use tools, maintain memory, and execute multi-step workflows with minimal human oversight.
The law firm analogy: A chatbot is like having access to Westlaw, you ask a question, you get information back.
An AI agent is like having a junior associate who can independently research case law, draft a memo, cross-reference statutes, flag issues for your review, file the document, and calendar the deadline, only asking for your approval at critical decision points.
What is an AI agent, really?
The five capabilities that define a true agent
Agentic Improvement
Law is a soft skill in the sense that it’s not “book me a plane ticket” type of agency. We’ll get into that later. But first, it’s important to see how agents are improving in general.
Agents-to-Agents Frictionless Future
Consumer: “Make a dinner reservation”
Agent: “Calls restaurant”
Restaurant agent sees an agent is calling: “Confirms availability”
Text pops up with confirmation
No people needed.
Claude Cowork
Let’s use this as an introduction to agents
Connect to a folder
Give it a task.
Come back to see what it built.
Real time.
Prompt on next slide….
Claude Cowork
Context Loading - Inn of Court Explanation Page Creation
I dictate to GPT’s built in microphone. And then I paste my prompt into the tool I need (Claude Cowork). I dictate as much as possible. Every time.
I have a meeting today with the American Inns of Court AI Virtual Inn of Court. I'd like the attendees to understand the history of how we got to where we are today. Please create a website that can be a resource for someone who is new to this organization that explains the journey from the Inns of Court in London, how that system was set up, places like the Middle Temple, Lincoln's Inn, and the English Inns of Court. From there, let's transition to the American Inns of Court, how they became to be, and how they first started, and the key similarities and differences to the English Inns. Give a brief history of the growth of the Inns of Court and how they have proliferated over the years across the United States. From there, we need to ground our conversation into the virtual Inn. For example, our hosts, Richard Herman, he has a past with the American Inns of Court. For example, the Richard K. Herman Technology Inn of Court, and then Kevin Brady and his role in the American Inns of Court. And then probably the least intuitive would be me. Ethan B. Holland. How do I connect to the Inns of Court? Give a little history of who my dad is, Randy Holland, and his participation in the history of the Inns of Court, his honorary bencher membership in Lincoln's Inn, and how I am his son and I am an artificial intelligence expert. Hopefully this website can create a nice thread for people who attend the meetings who can see the evolution from the English inns all the way through to the American inns, the growth and success and explosion over the years of the Inns of Court, and now for the first time ever, the virtual AI Inn of Court. Along the way, please provide links so that people can click on the websites, links, and learn more if they want to open new windows. This should be a wonderful resource for them.
Inn of Court Explanation Page - Completed
Parts of an agent
Agents are made of three core elements
The Model�The agent’s brain. Different models cost different amounts and do different things…�Think Golf carts v. Ferraris v. ATVs v. School Busses
The Harness�This is tells the agent what it can do, gives it access to to tools, limits scope, it’s the rules.
The Context�This is the prompt. And the documentation. The guidance and knowledge before it starts.
Claude Cowork- Agent Terms
Example of a Page Already Made (Gemini Image - CoWork Page)
Claude Skills: LinkedIn Profile Finder Demo
I made this for my daughter to find connections and learn LinkedIn
Output for each search:
This is saved as a “Markdown File” (I could share it with you!)
Claude Skills: LinkedIn Profile Finder Demo
Bonkers
Claude Skills: LinkedIn Profile Finder Demo
Claude Skills: LinkedIn Profile Finder Demo
Claude Skills: LinkedIn Profile Finder Demo
Claude Skills: LinkedIn Profile Finder Demo
Claude Skills: LinkedIn Profile Finder Demo
Claude wanted me to tell you this…LOL
Markdown Files
A term to know (a type of harness)
Markdown Files as Reusable Instructions
There are too many models to know
The landscape of major agent products is intimidating
OpenAI Products
gpt-5.4: Best intelligence at scale for agentic, coding, and professional workflows�gpt-5.4-pro: Version of GPT-5.4 that produces smarter and more precise responses�gpt-5.4-mini: Our strongest mini model yet for coding, computer use, and subagents�gpt-5.4-nano: Our cheapest GPT-5.4-class model for simple high-volume tasks�gpt-5-mini: Near-frontier intelligence for cost-sensitive, low-latency, high-volume workloads�gpt-5-nano: Fastest, most cost-efficient version of GPT-5�gpt-5: Previous intelligent reasoning model for coding and agentic tasks with configurable reasoning effort�gpt-4.1: Smartest non-reasoning model
Qwen, DeepSeek, Mistral, Zai, Grok, Claude, Sonnet, Opus, Haiku, Gemma, Gemini, and on and on and on….
There are too many models to know
OpenAI’s Page (we reviewed this on the fly)
Plus… consumer agents
The landscape of major agent products
Claude Cowork
Claude Code
OpenAI Agent
Gemini Deep Research
Agents with “computer use”
Claude For Chrome, OpenAI Agent, Perplexity…��Prompt: Can you track down the AI Virtual Inn of Court and send an email to someone to request information on how to join? Use my Gmail.
How agents work together
Multi-agent orchestration using a legal team metaphor
The "digital law firm" concept
Just as no single lawyer handles every aspect of a complex case, no single AI agent handles every aspect of a complex workflow. Multi-agent orchestration coordinates multiple specialized agents working together. Each focused on what it does best.
Handoffs: One agent transfers control to another when it encounters a task outside its expertise. The first agent packages up all context and passes it along. This is like a triage nurse who takes your symptoms, then transfers you, with notes, to the right specialist.
Swarms: A model with no central controller. Agents operate as autonomous peers making local decisions based on shared state. OpenAI popularized this with its experimental Swarm framework (October 2024). Imagine a colony of ants: no individual ant knows the full plan, but the colony accomplishes complex goals through simple local rules.
Orchestration patterns (from simplest to most complex):
Swarm/Mesh: Fully decentralized peer coordination
How agents work together
Examples
xAI’s SuperGrok Heavy ($300/month) uses a true multi-agent architecture internally. One agent plans, another checks the plan, a third writes the output, all working in parallel. It scored 50.7% on Humanity's Last Exam, the first AI to break 50% (summer 2025)
Legal application example: Lexis+ AI's "Protégé" system deploys four specialized agents simultaneously…… all coordinating to answer a single complex legal question.
Big question: Are they wrappers?
Are enterprise AI law tools “wrappers"
Big question: Are they wrappers?
The market crushed Thomson Reuters
Huge gaps (??) in law performance
Let’s scrap THIS and look at hallucinations…
X
Huge gaps (for me) in law performance
I had Gemini Pro Deep Research, GPT Pro, and Claude Deep Research Sonnet 4.6 each run a report on AI legal tool hallucinations. In a rare twist, lawyers defended the tools as not hallucinating as much as the reports suggest. My instinct is the data is old and there is not enough benchmarking to get a true read on performance….
See the reporting here:
https://ethanbholland.com/2026/04/21/ai-hallucinations-for-lawyers/
It would be neat to get the AI Inn to build a benchmark!
Discussion on law benchmarks (on the fly)
Discussion on law benchmarks (on the fly)
Discussion on benchmarks (on the fly)
Discussion on benchmarks (on the fly)
On the fly chat
Enterprise -> Frontier -> Open Source?
On the fly discussion
On the fly discussion - OpenSource lag
On the fly discussion - OpenSource
On the fly discussion - OpenSource
OpenSource: OpenClaw
OpenClaw is a self-hosted, open-source AI agent platform created by Austrian developer Peter Steinberger (founder of PSPDFKit). Originally launched November 2025 as "Clawdbot" (a play on Claude), it was renamed after Anthropic trademark complaints. It became the most-starred repository in GitHub history — over 285,000 stars by March 2026. Jensen Huang called it "probably the most important software ever released."
OpenClaw runs as a Node.js gateway connecting 20+ messaging platforms (WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Teams) to any LLM. It features a heartbeat scheduler for autonomous operation without being prompted. Notable use case: one developer's OpenClaw agent negotiated $4,200 off a car purchase via email while he slept. Another user had it draft a legal rebuttal to an insurance claim. It's MIT-licensed and free, though security researchers have found prompt injection vulnerabilities.
Runs locally - download the “harness”…. Pick a model. It goes BANANAS.
OpenSource: Nous Research Hermes
Nous Research is an independent US AI lab out of New York that builds high-quality open-weight models. Hermes 4 (August 2025) introduced "hybrid reasoning" models that can toggle between standard responses and explicit chain-of-thought reasoning.
Hermes Agent (March 2026) is their open-source persistent personal AI agent with self-improving capabilities, multi-model routing, and 40+ built-in tools. Their latest Hermes 4 35B model (April 2026) runs on a single consumer GPU and achieves roughly 70% of Claude Sonnet 4.6's agent quality at zero cost. All MIT-licensed.
The open-source framework landscape
LangGraph (LangChain): The leading framework. Used by Uber and Klarna (which saved $60 million annually). 34.5 million downloads in 2025.
CrewAI: The most accessible. Define a "crew" of role-playing agents (Researcher, Writer, Reviewer). 44,300 GitHub stars. Can prototype in 2–4 hours.
LlamaIndex: Best for document-heavy applications. legal research, contract review, compliance. Handles 90+ file types.
Microsoft Agent Framework: The enterprise merger of AutoGen and Semantic Kernel.
Model Context Protocol (MCP): The crucial infrastructure standard, created by Anthropic in November 2024 and donated to the Linux Foundation in December 2025. MCP is "USB-C for AI" — a universal adapter that lets any AI connect to any tool through one standardized protocol. Adopted by ChatGPT, Gemini, Microsoft Copilot, and thousands of community-built connectors.
The total open-source agent framework market saw 34.5 million downloads in 2025. A 340% increase from the prior year.
Self Service
Even WITHOUT agents (it’s blurry).. individuals are using AI tools with success.
AI Satisfaction: Less Friction, Better Outcomes
Self Service
AI Is Removing the Bottom Rungs
Real-World Examples
Dialogue: What responsibilities do consumer tools have if they provide legal advice?
On the fly discussion - Self-service?
Self-Service Cases: Built in 5-10 minutes