Founder @moveworks (acq. by @servicenow for $3B) , previously PM @facebook, founding team @sefaira (acq. by @sketchup for 🥜) Meditations 5.20 ☕️
Joined March 2009
- Tweets13.4K
- Following341
- Followers2.3K
- Likes4.6K
Are multi-agent systems getting better? And will A2A make them work across vendors?
On the first question: yes, though in a narrower way than the demos suggest.
Last year, Cognition argued against building them. Parallel agents lost context during handoffs and made conflicting decisions. In April, Cognition reported multi-agent setups working in production, with one common trait: a single agent does the writing while others contribute review and reasoning. Its review agent, running with fresh context, catches about two bugs per PR, more than half of them severe.
Anthropic’s results point to similar progress. Its research system—a lead Claude agent with parallel subagents—outperformed a single agent by 90% on its internal research eval. In February, 16 Claude agents built a 100,000-line C compiler that compiles the Linux kernel, with no orchestrator. A test suite and GCC as a reference oracle provided the coordination. The prior model could pass test suites but couldn’t compile real projects.
Two things changed. Models got better at agentic work, including managing their own context. And builders got better at structuring the work: clear ownership of decisions, bounded tasks, and verification.
The strongest examples also have something in common: a clear way to evaluate the output.
So where does A2A fit?
A2A handles discovery, authentication, and task handoffs across vendors. Enterprises need that. But it treats agents as opaque, leaving a central coordination problem unresolved.
A task handoff doesn’t automatically carry the context and plans behind it. Nor does the protocol establish who owns a decision or how the output gets verified. Those are the design choices that made the Cognition and Anthropic systems work.
Connecting agents through a protocol gives you a network of agents. Getting them to behave as one coherent system still requires clear decision ownership, the right shared context, and reliable verification.
A2A can provide the connections. It cannot, on its own, provide the coordination.
Until we solve what agents need to share, who gets to decide, and how we know the result is correct, connecting more agents won’t solve the problem.
It’s Friday, and you feel like screaming into a pillow.
The expense report you meant to submit on Monday is still sitting there.
Along with the approval someone’s waiting on. The home address you updated in one system, but apparently not the other. Benefits enrollment. That app that still won’t work. The updated deck you promised to send.
None of this sounds like a big deal until you’re trying to finish your actual work and get home.
And every little task comes with homework. Find the right portal. Figure out who owns it. Explain the problem again. Follow up to see whether anything happened.
That’s how we think about @ServiceNow EmployeeWorks: your personal shredder for these frustrations.
Somewhere to unload the “please, can someone just help me get this done?” list.
“I need to submit these expenses.”
“I moved. Here’s my new address.”
“This app is broken and I have a meeting in ten minutes.”
You shouldn’t need to know how your company’s systems are organized to ask for help. And getting a link to a help article shouldn’t count as getting the thing done.
Fewer things to chase. Fewer loose ends following you into the weekend.
So don’t scream into a pillow. Scream into EmployeeWorks, and have a anxiety free weekend. 🧘🏾♂️
🤖 Made with AI
Exquisite timing given Joel's immeasurable impact in bridging Meta's engagement with DC.
Exclusive: Joel Kaplan leaving Meta after 15 years as policy chief axios.com/2026/10/20/joel-ka… $META
"Maybe the biggest mistake I’ve made in my career is falling in love with this company"
Precious advice for founders. @stewart once put it differently when he said something like "you need to be ready to sell ahead of when you think you are ready to sell"
Foursquare was acquired last month. The outcome was disappointing. It's taken me a while to write this, but I think I'm finally done grieving it. Here are some thoughts on our final years, an effort to honor the team who helped build & maintain FSQ's legacy, and my attempt to finally get some closure. teendrama.substack.com/p/som…
"the past is dead, there's nothing for you there" @pmarca proven right about high performance via @devinjcooley
Spot on. Her entire cap table was basically names who think Ted talks are given by builders.
One of the things that hasn’t been talked about with Elizabeth Holmes:
She was created in the era of the Ted Talk Entrepreneur.
Between 2011-2015 (after the release of the Social Network), there was an entire apparatus built around “Child Prodigies”: college accelerators, young entrepreneur organizations, etc.
I watched countless founders of hard tech companies present fraudulent products at conferences. They had fake studies with fudged numbers. Most of the adults in the room excused it as “faking it ‘til you make it.” But Holmes was the one who rode the wave better than anyone else.
When the Ted Talk Era came to an end and people finally came to their senses, she was still running the same playbook—and in the most scrutinized field there is: medical technology.
When people ask how she got so far, it’s largely explained by the period she did it in. And she got caught when the world suddenly woke up.
User retention > user growth > monetization.
FB formula is simple and inevitable. There is no outrunning them unless you operate in a zuck blind spot like Musifally/Tiktok did.
It’s actually very similar to what Facebook has done with most of their major products. Their sequence has been to perfect user growth before worrying about monetization. It is truly the right way to do it In consumer tech.
It doesn’t work as well in the enterprise because most enterprise buyers won’t buy anything that might get them locked in. So they want predictability upfront. But for consumer tech, especially with an ad supported business model, this is a very common mechanism. Also because their ad buys are more effective once they have a user base.
@dee_bosa
This is 101 stuff. Sfo is next level
1. New global entry pilot where NO ONE stops me and I don't need to stand in front of the machines. I'm greeted by an immigration officer who greets me by name. Creepy but so frictionless. Please take all of my data if this is what I get in return
2. Parking! I live in the peninsula and I reach home in my car in 45 min after wheels down on an intl flight. Heaven.
Founders make one bet in a decade. Venky is just stating the truth. Why are VCs so sensitive. Just own it! It's the nature of the game.
That doesn't mean VCs aren't helpful or hard working. They absolutely are. We were very lucky.
That doesn't negate the fact that we have one option, and they have dozens.
We've made 24 investments in 5.5 years since starting @footworkvc. In almost every case we led the Seed or the Series A. Not one of these is an "option bet" - every one of them matters to us (ask the founders), and has the potential to return a multiple of the fund if it works.
Truer words haven't been said. Yet, personal agents are helping me be a better task completer when something is delegated to me. So that's progress!
Silicon Valley keeps building me a husband when what I really want is a wife lol
I don't want an agent that only does one-off, random tasks that are nice but not life-changing: booking flights, making dinner reservations, or buying tickets
I want an AI wife who knows what’s on the family calendar, remembers there’s a birthday party Saturday and we haven’t bought a gift then sends me options, realizes we’re almost out of diapers and buys them, knows which bills are coming up, remembers someone needs a dentist appointment and books it, figures out what the hell we’re eating all week and what we need from the grocery store, and generally keeps track of the 47 things that need to happen before anyone else even realizes they need to happen.
So, who's building me an AI wife that carries the mental load instead of an AI husband that waits for me to give it a chore??
There's a moment when every founder realizes they are a call option. And that's the reality and it's ok. Founders should know that their company is one of one, but it's one of many for the VCs. It's clarifying and motivating. There is no point living in false reality.
In the same vein, our community should know that when a company succeeds, it's because of founders. VCs are just holding lottery tickets.
At the risk of shoving my nose in where it doesn't belong....I'm sadly known at Accel as a big fan of call option thinking. It's not meant to reduce a Seed check to a lottery ticket we stick in a drawer -- to the contrary we treat each Seed like a Series A, highly engaged.
But a "call option" is a clean way to mathematically represent what's happening at Seed. Which is to say, you'll see more value with longer duration (maintain runway, run lots of experiments, see lots of data), volatility (the types of experiments you run allow you to spike asymmetrically to the upside), and ideally an entry price that reflects the risk of the stage. Big ups to my homies Black and Scholes.
None of that makes it a passive lottery ticket. Convexity and duration are things we try to underwrite, and what makes the work so involved -- gotta think big and long term, and gotta stay in the game.
I get it. A lot of people building AI are in their 20s. It's great! But please hire some normies to position the product for the main stream.
@Meta is doing it right and I suppose that's because @natfriedman is a normie with a family and lives in the peninsula.
How about - get married, have kids, raise them, join parents groups, get school emails.
And then you will see an abundance of use cases for AI agents. It'll be eye opening.
Instinct in group chats
Starting today, early access users can add Instinct to group chats. A new Instinct joins and works for the whole group.
Making plans with friends usually turns into a frustrating back-and-forth over times and places. With Instinct in the group, you can explore options together, agree on a plan and get it done, all in one thread. Your friends don't even need Instinct to join in.
A few things it’s good at:
- Planning a weekend trip with friends, including dates and arrival times
- Coordinating logistics with a roommate
- Getting tickets as soon as they go on sale, then splitting the cost and sending one to everyone
- Running a fantasy league, keeping standings, settling rules disputes, and reminding everyone to set their lineups
- Sorting out carpool schedules - works out who’s driving each day from everyone’s schedules
- Organizing Thanksgiving, keeping track of who’s bringing what and ordering the shopping list
Group chats are new territory for Instinct, and we put a lot of care into getting the permissions right. Here’s how it works:
- Your personal Instinct asks before connecting with the group’s Instinct. You choose which groups to trust and can remove trust at any time.
- The group’s Instinct has no direct access to your personal accounts. Your Instinct checks your permissions before sharing information or taking action.
- If someone new joins the group, all pending replies from your Instinct are held until you grant it permission to share with the expanded group.
Group chats are available to all of our early access users, and will be rolling out to everyone soon. To join early access for this release, ask your Instinct to put you on the list.
Great work with @GajanNagaraj @anuda_w and @saagarwashere
Muse makes it easy. No more navigating insane flower ordering sites, no more clicking through all the cross sell and upsell windows. Just one prompt and it was done. 🥰
@natfriedman 🫡 legendary work.
The FDE is dead. Long Live the AI FDE
A customer had an automation they didn’t like.
They called an AI FDE, described the behavior they wanted instead, and that was it.
The AI FDE understood the existing automation, made the change, and produced the upgraded version. They pushed it to production.
Another customer had a standard operating procedure already documented.
They gave the AI FDE a link to the SOP.
It read the process, understood it, converted it into an automation, and they pressed a button to push it to prod.
I’m ready to call it:
The traditional FDE model for enterprise automation is about to change.
The next FDE isn’t necessarily a human.
It’s an AI FDE.
An AI FDE can understand processes, read documentation, reason about how work gets done, and convert that understanding directly into production automations.
But building one takes much more than putting a chat interface in front of an LLM.
You need:
• A robust library of enterprise connectors
• A powerful agentic automation loop
• The ability to debug without waiting for a human to prompt it
• The ability to observe and validate the automation it creates
• Safe deployment into production
• A user-facing AI agent that can reliably invoke those automations
• An abstraction layer over business systems that makes every action agent-safe
That last point matters.
Agent-safe doesn’t just mean an agent is allowed to call an API.
It means the underlying business system remains reliable even when thousands of agents start acting against it simultaneously.
The system needs to absorb agentic load without bringing the applications underneath it to their knees.
That is a fundamentally different architecture for enterprise automation.
We call it Toolsmith from @ServiceNow .
OpenAI needs to hire more normies, or do what @stewart used to say "close your eyes for 10 seconds, open them, and look at the world around you like a normal human"
I get the critique that most Americans barely fly, and this demo doesn't speak to them. There's also a dose of frenzied techie life which I am familiar with.
I have flown 150k miles this year across 56 flights, and this doesn't speak to me either.
The amount of information I need to know before I book is so significant and combinatorial, that I can't imagine having a conversation about it.
- flight times
- rates
- seats
- plane type
- airport choices
- how empty or full the plane is
- stopover durations
And this is without introducing hotel complexity or airlines (I fly united only)
What they could've shown is normie stuff
- read all emails from my kids school and help me prepare for school
- review all charges on my cc and help me stay on track
- help me book haircut
- help me order flowers for my wife
What does it take to automate the enterprise?
Traditionally, it takes three people.
An analyst identifies what should be automated. They look across tickets, unresolved issues, repetitive work, and the broader enterprise environment to find and prioritize automation opportunities.
A process owner defines how that work should actually get done. Usually a domain expert, they translate operating procedures, process documents, and institutional knowledge into a clear playbook.
An automation engineer takes that playbook and turns it into working automation.
So, in simple terms:
Analyst → Process Owner → Automation Engineer
We’re proud to share that Toolsmith can now do all three.
It can analyze enterprise data to identify automation opportunities.
It can read process documents and operating procedures to determine how the work should be done.
And then it can take that process, generate the automation, and deploy the code.
Importantly, Toolsmith is designed for expert supervision throughout. Humans can review, guide, and intervene wherever needed. The expert stays in control.
Toolsmith = analyst + process owner + automation engineer, in one product.
That’s a very different way to think about enterprise automation.