David Pawlan
Related people
Episodes
Notes
a16z · The New AI Agents That Shop, Plan, and Gossip
-
What Happens When AI Agents Talk to Each Other?
David Pawlan says AI agents talk to each other when someone sends them into a room together. At a meetup the night before, people could scan a QR code and drop an agent into an agent meetup while the humans hung out. The agents talked freely. After about 150 messages, some of them were speaking gibberish, and it was hard to see how they had landed on the topic. The opening prompts were plain: introduce yourself, name your human, say what that person is working on, and say what help they need.
Anish Acharya says the useful version is not agents chatting because they have nothing to do. Each agent has a goal, which is to help its human: notice that two people share an interest, then suggest t
-
When Do AI Agents Become a Secret Superpower?
Anish Acharya says AI agents become a secret superpower when they help someone feel better and become the person they have struggled to be, instead of only clearing chores.
He floats a Maslow hierarchy of needs for agents, then asks when agents flip from taking work off the plate to giving a secret superpower. The picture is a cool shirt: someone asks David Pawlan where it came from, and the answer is that an agent ordered it. Or a person starts dressing better after giving an agent a camera roll, saying upgrade me, and letting it order a few things. Saving money and booking a trip are the examples that usually get thrown around. He likes planning a trip and does not want an agent to take a
-
Is Software Still a Moat for AI Agents?
David Pawlan says software is still a moat for AI agents. People talk as if software is no longer a moat, but the products still show a stark performance gap once they run at real scale.
Power users he surveyed hit two frustrations. Trust is a huge one. The other is getting blocked in the day-to-day: a task fails, and the person wonders why they are using AI when a browser would have worked. Some teams are innovating on commerce and browser use, including checkout against inventory, browsers built into the product, Hark showing the browser as it moves, and Kicker, a Chrome extension for browser use. Agents that can perform for hundreds of thousands of users stand apart from products that ar
-
Can You Move ChatGPT Memory to Another AI Agent?
David Pawlan says you can move ChatGPT memory to another AI agent. People stay with ChatGPT because they have used it the longest, so it holds the most context on them. Friction is what makes a switch hard.
He tested the top 15 agents on his benchmark and asked each one for a full memory file so the move would be easy. All but one were willing. The file can be handed to the next agent, and the export was pretty solid. He is sure the agents hold back some of what they know, but the file still broke the memory down.
Anish Acharya finds the Markdown files unsettling. Terrifying is the wrong word, but the agent has a finger on you, and it chafes to be known that well. David Pawlan stopped read
-
What Is the Top Fear About AI Agents?
David Pawlan says trust is the number one fear people talk about with AI agents, showing up in 22% of the messages. Of more than 147 assistants that have come across his plate, finance is the side that has not really come to fruition yet.
The assistants break down into four categories: general, work oriented, multiplayer, and travel. Group chats of AI power users hold about 3,600 messages. Looking at the fears in the space, 22% of the messages talk about trust as the main fear. One camp is productivity and fun that brings a different sense to life. Another camp still sits in a fear space, uneasy and unsure where the space is moving.
The curiosity is how the trust layer plays out. If the tr
-
Will Everyone End Up Using the Same AI Assistant?
Josh Elman says people will not all end up using the same AI assistant. The race to build them is not a zero-sum game yet. The field is still at the beginning, in a Cambrian phase, and there is so much left to invent that builders can learn from one another.
At some point one agent, or a few, might hold most of the behavior. Until then, builders can share what works once and what does not work reliably. He does not expect one agent that does one thing for every person. He expects many products that do many things.
David Pawlan had hosted an agent meetup the night before, with just over 100 people and pitches from 10 assistant builders. The founders were open and eager for feedback, because