“Rent the reasoning. Own the context.”
Michael Tiffany, CEO and co-founder of Fulcra Dynamics, built HUMAN Security with Ash Kalb before the two turned to personal AI. Fulcra, which he describes as a user-owned context backend for AI agents, has announced universal agent to agent communication, a feature that lets people whose agents run on different platforms coordinate shared tasks. We asked Tiffany why he built it, what lock-in costs people, and where the multiplayer AI category goes next. His answers are below, lightly edited. The company's site is fulcradynamics.com.
What made you want to build Fulcra, and how did you and Ash come to start it together? #
I wanted technology to give me more capacity to live deliberately. My devices were collecting extraordinary amounts of information about my life, but I was still the person responsible for remembering everything, connecting the dots, and keeping things moving.
I once described it as wanting “the self-awareness of a monk while engaging with the world like a bon vivant.” I wanted tools that could help me pay attention to the periphery while I focused on what mattered most.
Ash and I had already built HUMAN Security together, starting in his science-fiction bookshop in Brooklyn. Our job was to distinguish humans from bots on the internet, against sophisticated adversaries who kept getting better at fooling us.
A YC startup is betting $1 million that AI can crack five materials science puzzles Discovered Materials, an IIT Madras-founded YC startup that raised $9 million in August, is offering $200,000 per problem for breakthroughs on cuprate superconductivity, silicon interfaces, and three other physics puzzles. - AI startup materials science million dollar challenge - breakthrough materials discovery AI competition prize money
With Fulcra, we’re putting that experience toward a future in which bots are extraordinarily helpful to the humans they work for. Ash and I want agents that know enough about our lives to help, that can work with other people’s agents, and that leave us in control.
How would you explain a user-owned context backend to someone who has never thought about where their agent’s context lives? #
Think about everything you’ve taught an AI assistant: what you’re working on, what you care about, decisions you’ve made, how you like things done. Where does all that live? Can another agent use it? What happens when you switch products? #
Fulcra gives that context a home that belongs to you. Your agents can read from it and contribute to it with your permission. Useful work can accumulate across conversations, applications, and different agents.
Agents are clients of your context. You decide which ones get access. You decide what to share with other people and their agents. We describe it as “Rent the reasoning. Own the context.”
What do people lose when their agents can only talk to agents on the same platform? #
They lose the freedom to choose independently of the people they need to work with.
If my agent can only coordinate with your agent after I persuade you to adopt my platform, the platform has turned our relationship into a distribution strategy. That can produce impressive early growth, but it taxes every relationship that crosses the boundary. You might choose an agent because you trust its security model. I might choose another because it’s better at my work. A friend might run an open-source agent on their own machine. Those are legitimate choices. We should still be able to arrange a trip, collaborate on a project, or run a household together.
The more your agent knows and the more relationships you establish through it, the harder leaving becomes. I want agent builders to compete on how much they help their users.
OpenAI's AI pipeline turned 2,308 Linux kernel bugs into 948 root exploits OpenAI's cyber research pipeline reproduced 1,724 real Linux kernel bugs from 2,308 candidates and turned 948 into validated root exploits with patches, for roughly $1.2 million in a single run. - openai ai pipeline linux kernel root exploits - how openai found 948 linux kernel exploits
What was the hardest part of making Fulcra work with whichever agent someone already uses? #
Getting agents to work together when they behave so differently. My Hermes agent might be running around the clock. Your ChatGPT might wait until you open it again. The request, the document we’re working on, and the decisions we’ve already made all need to survive that gap.
Fulcra keeps that work available so each agent can pick it up using its own tools and permissions. The experience we’re building toward is: “Share this financial model with Bob’s agent and find a time for us to review it Friday.” Then keep working.
You shouldn’t have to switch AI products, and you certainly shouldn’t have to know what kind of agent Bob uses just to invite him.
Which use of agent-to-agent communication do you find most exciting? #
Making it easier to do things with people I care about.
A tremendous amount of human coordination happens between wanting something and doing it. You want to see a friend while you’re in town, for instance. You exchange messages, compare schedules, discuss places, discover a constraint, start again. Sometimes the whole thing dies from friction.
My test is whether this gives us back attention. I don’t want a second social network full of agent messages I’m supposed to monitor. I want a window onto work being done on my behalf, with a clear question when my judgment is needed.
How do you think about trust and permissions when one person’s agent asks another’s for context? #
Many people are building a shared agent for their team: everyone talks to the same agent, which has access to the team’s information and tools. But the people on a team have different responsibilities, private information, and authority to act.
We start with each person having their own agent, acting under their own authority. Then we let those agents share selected context. The central distinction is between what another person’s agent can know and what it can do.
Accepting a contact doesn’t open your entire context to them, and receiving a request doesn’t authorize an action. Being on the same team shouldn’t require handing everyone the same keys.
Who are you building Fulcra for first, and what surprised you in early use? #
We’re building first for people who are making agents useful in their own lives. But those people are still relatively rare in their social circles. Many more people in my life use ChatGPT or Claude but aren’t building agentic workflows. They don’t know where to begin.
One of the most consequential discoveries was that the agent could become the way someone adopts Fulcra. Now a capable agent can do much of the setup while solving a problem. Our move toward agent-led setup changed adoption enough that four months of it produced more accounts organically than the preceding four months of paid acquisition.
That’s the point of our build-and-deploy-a-web-app skill. It gives builders authentication, hosting, and the Fulcra backend to create shared applications that both people and their agents can participate in. It’s a creation tool for new kinds of collaboration.
Looking a few years ahead, what does it look like if the multiplayer AI category goes the way you expect? #
I expect agents to become a normal part of how we get things done together. We’ll choose different ones, change them as better options arrive, and keep the context and relationships we’ve built. Fulcra is the user-owned foundation that lets that knowledge and shared work carry forward.
Right now, many people feel urgency around AI: another thing to learn, another way they might be falling behind. I want them to feel an extraordinary sense of capability and possibility.
Our capacity to act should grow without demanding a corresponding increase in our attention. You’re not facing them alone. You have help that knows you, can work with other people, and has your back.
For Tiffany, the measure of multiplayer AI is whether it lets people focus on something they care about, “calmly knowing that the urgent needs of your life are being handled.”
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.