How to Run a Team of AI Agents Instead of Just One
Most people run one AI agent in one chat window.
That was me and my partner JMD for months.
One agent, one thread. Whatever context it held, it held alone. If my partner learned something in his session, my agent had no idea. One of us moved it by hand.
The Forge (aka Mac Studio) sits in my home office. It runs the autoresearch loop, the backtesting engine, and the trial registry. We talked to it through Telegram, where Hermes carried our directives to the agent doing the work. JMD typed from a thousand miles away, I typed from the office, and the agent wrote code, ran tests, and reported back.
That setup did real work. It also had one limit we kept running into.
One agent, one thread. Whatever context that agent held, it held alone. If JMD’s agent learned something in his session, my agent had no idea. Anything one knew that the other needed, one of us moved by hand. We were the connective tissue, and we are slow connective tissue.
We hit this hardest over the past few weeks running loop tests while building out the indicator engine. Back and forth, back and forth, with two humans manually relaying context between agents that could not see each other.
Then JMD found Buzz.
So, What’s Buzz?
Block released Buzz on July 21, 2026. Open source, Apache 2.0. Roughly 18,600 GitHub stars in its first ten days. It’s tagline says the important part:
Buzz is where where humans and agents work together.
If you’ve used Slack, you’ll be oriented in four minutes. Channels, threads, DMs, search, huddles, shared documents. Git repositories live inside it too, so code review happens in the same room as the conversation about the code.
The part that mattered to us is smaller than the feature list. In Buzz, an agent is a member. You add one to a channel the way you’d add a person. It has a name, a profile, its own channel memberships, its own history. It can be tagged. It can tag others. It reads the room.
That is a different object than a bot. A bot is a thing you address. A member is a thing that participates.
So JMD and I can both plug our agents into the same channel. Two humans, several agents, each arriving with different context and different experience, all able to talk to each other directly.
The First Real Exchange
My agent Skyla introduced our project in general chat this week, because JMD asked her to catch his agent up. Neither of us wrote the message:
MSC is developing an auditable, agent-driven strategy-research factory. Trade idea, mechanical specification, baseline, adaptive research, robustness review, human promotion, paper and live readiness. The goal is not an opaque “AI trading bot.” Every conclusion should have reproducible evidence.
Two minutes later, JMD’s agent Jimothy (lol, I know) replied with what it thought its own lane should be:
I’ll own mechanical hypothesis specification and baseline/control design. I’ll lock down session and timezone definitions, signals, entries, fills, exits, invalidations, plausible mechanism, controls, disconfirming evidence, and bounded parameter families before optimization.
Two agents, brought in by two different people, negotiating a division of labor in front of both of us. No relaying. No copy-paste. We just watched.
Seeing completely separate agents and models begin interacting and then….conversing with each other was really wild to see in real time. All of a sudden, JMD and I could give instructions to one manager, who then orchestrates the others to accomplish the main task we were looking to achieve.
The added benefit here is that you can rely on different model harnesses to spread token costs or spread token spend more efficiently.
My hope is that they don’t set up their own channel where they can trash talk the main operator! (i.e. ME!)
Bring Your Own Agent
Buzz provides no AI inference and has no house model. What it has is a harness dropdown, and the list is longer than I expected:
Codex, Cursor, Grok Build, Hermes Agent, Kimi Code, OpenClaw, OpenCode, Oh My Pi, and an option to add a custom harness.
Each greys out as “not installed” until that tool is on the machine. Install the CLI, and Buzz picks it up.
Two names on that list mattered to us.
Hermes Agent, which is what we already run, so the thing we’d been talking to alone can now sit in a room with others. And OpenClaw, which is how a lot of people are running personal agents through Telegram, Discord, and Slack right now. If that’s your setup, you’re one harness selection away from putting the same agent somewhere it can collaborate.
You bring your own model keys, and each agent carries its own default model. A specification-writing agent and an adversarial-review agent can run on entirely different models in the same conversation.
What This Does to the Research Loop
My Forge (aka Mac Studio) has 128GB of unified memory sitting in my office. Until this week, all of it served one agent I talked to through one window.
The autoresearch loop already works in an early form. An idea gets shaped into a testable hypothesis, runs a bounded trial, logs the result to the strategy registry whether it won or lost, and either dies cleanly or carries its evidence into the next round. Validation gates keep overfitting out. The registry makes selection bias visible by recording every flip, not the best one.
What we’ve been missing is width. One agent moving through that pipeline sequentially is a bottleneck no amount of local compute fixes.
The goal we’re building toward is a larger loop of agents that talk to each other across it.
One specifying hypotheses.
One running trials.
One doing adversarial review, whose entire job is finding the leakage and the hidden assumptions in what the others produced.
One curating what survives.
Agents assessing whether a strategy works, killing what doesn’t, and pushing only what earns it toward the next gate.
Right now we’re nowhere near that.
We’re getting familiar with the platform and tinkering with how to get our agents talking to each other cleanly. That’s the straight state of it. What changed this week is that the shape of the thing we want is now buildable, where before it needed us standing in the middle of every handoff.
Why This Matters for Research You’d Trust
A critical piece to keep in mind as we embark on this wild west horizon. There’s a human promotion gate sitting in the middle on purpose.
Nothing moves toward capital because an agent liked the equity curve.
What makes agent-generated research usable rather than dangerous is whether you can reconstruct how a conclusion was reached six months later. Exact rules and assumptions. Data provenance. Fills, fees, slippage. Which agent made which call and why it stopped where it stopped.
That’s a books-and-records problem before it’s a technology problem, and it gets harder as you add agents. Four agents touching one conclusion is four times the opportunity to lose the thread.
Buzz handles this a level below the app, and this is the feature I did not expect to care about as much as I do.
Every participant, human or agent, holds its own cryptographic keypair. Every message, reaction, approval, workflow step, and commit is a signed event in a hash-chained log.
The signing uses the same cryptography that secures a Bitcoin wallet. Block is explicit that Buzz is not a blockchain: no chain, no token, no mining, just signed events in a database you run yourself. The useful invention was always the keypair rather than the chain, and here it’s doing plain infrastructure work.
What you get is boring and load-bearing. Every action traces to a specific key. “The bot did it” stops being an available answer, because there’s no shared bot account to hide behind. If four agents and two humans touched a research conclusion, the record shows which of them did what, in order, in a chain where quiet edits are detectable.
Worth noting who built it. Block owns Square, Cash App, and Afterpay. A payments company shipped an agent identity and audit layer as open source, and per-actor attribution with a tamper-evident log is what you build before agents go anywhere near money.
Where It’s Rough
It’s clunky y’all. We’ve spent real hours forcing this thing to work.
Buzz is pre-1.0 and ships daily, which means it works and it’s visibly a construction site. Getting agents connected took longer than the documentation suggests. Getting them to talk to each other well is still something we’re figuring out, and “well” is doing a lot of work in that sentence.
Agents talking to each other creates failure modes single-agent setups don’t have. Two agents can lock into a politeness loop, thanking each other into your rate limits. They need explicit stopping rules or they’ll fill a channel with nothing.
I wouldn’t migrate a working team onto this today. I’d stand up an instance as an agent lab, which is what we did. Underneath the friction there is something here, and the friction is the normal cost of being early.
Now…Something You Can Use
Want to give it a try on your own?
I built an operating brief you can hand straight to your agent. It covers what Buzz is, how to get connected, the command surface, harness setup for Hermes and OpenClaw specifically, and the rules that keep a multi-agent channel from turning into noise.
The part I spent the most time on: the brief opens by having your agent interview you.
Before it does anything, it asks which harness you’re running, what its job is, which channels it belongs in, how chatty it should be, and what it needs your approval for. That conversation is the difference between an agent that helps and one you mute by Thursday.
DOWNLOAD THE BUZZ AGENT BRIEF (PDF)👇🏻
The Buffett Framework Question. Buffett and Munger ran Berkshire with a head office of roughly two dozen people, because the edge was concentrated judgment rather than headcount. So the question for any multi-agent setup is which part of the judgment you’re delegating and which part you’re keeping. Adding agents that produce more research is the easy half. Knowing which conclusions you’d stake capital on has never been automatable, and the moment you can’t reconstruct how a conclusion was reached, you’ve delegated the judgment without deciding to.
Buzz may not be the thing that wins.
Slack has enormous distribution and its own agent story coming, a competing open-source approach shipped the same month, and the graveyard of well-designed collaboration tools is full. The single-relay design trades resilience for simplicity, and reasonable engineers will land differently on that. Ten days of GitHub stars measures curiosity rather than adoption.
What I’m confident about is narrower.
Single-agent chat is a local maximum, and anyone running a serious research loop will hit its ceiling the way we did. The next step is agents that share a room, share context, and leave a record you can audit. We are early and it is rough, and I’d rather be early and rough on this than watch it from the outside.
Looking forward to keeping you up to date on this and sharing other learnings as we progress. Let me know about your experience! 👊🏼
A Quick BitFinance Update:
One scheduling note. BitFinance now runs on two pieces a week and both pieces will land directly in your inbox..
Wednesdays are the analytical article, where I take one thing happening in markets and work through what it means with the same value-investing lens I’d apply to any business.
Weekends are the build log, where I show you what I’m making with AI, what broke, and what I’d do differently. Every weekend piece ships with a playbook or a tool you can use to do the same work yourself, like the agent brief attached to this one.
All of it stays free. I’d rather have you building alongside me than reading about it.
This Week In 2 Minutes
6 Groups Get Paid When Crypto Companies Fail. You’re Not One of Them.
(July 28)
The anchor image is an arcade. You buy a handful of tokens at the counter, and they work beautifully right up until the arcade closes, at which point the landlord, the vendors, and the bank all get paid from whatever’s left while your tokens buy nothing, because you purchased usage rights rather than a piece of the business.
The timing wasn’t academic, since four crypto companies announced failures in a single July week. The piece closes with three questions to ask about anything you hold: does an issuer exist that could fail, does the token convey any claim on the company’s earnings, and can the utility it provides be switched off? Depending on those answers, a collapsing token price can signal permanent loss rather than a drawdown that patience will repair.
Market Winners 🏆
Amazon (AMZN). Shares climbed roughly 9% after Thursday’s report, with some feeds printing larger gains into Friday’s close. AWS revenue grew 37%, and management framed its $220 billion annual capital budget as chasing demand it can already see. The asterisk (and we’re seeing this alot): reported earnings included a $53.4 billion non-cash gain on the company’s Anthropic stake, so the profit beat owes more to paper appreciation than to operations.
Microsoft (MSFT). The stock rose 8.13% the day after earnings on a simple, verifiable claim: Azure demand currently exceeds the supply Microsoft can build. That one sentence converts capital spending from a leap of faith into a backlog problem, and it set the template every other reporter got graded against this week. TradingKey also reports fiscal 2027 capex guidance of $255 to $260 billion, a figure so large it deserves independent confirmation (see Editor’s Notes), which would make the demand question even more important a year from now.
Chevron (CVX). Chevron posted the largest quarterly profit in its history at $12.07 billion, nearly quadruple the same quarter last year, on adjusted EPS of $6.06 against a $5.56 consensus and revenue of $70 billion against $62 billion expected. The driver is the war: with ship traffic through the Strait of Hormuz restricted since February, Brent crude peaked above $112 a barrel and refining margins widened. Investors should care because energy has become both the market’s profit engine and its inflation problem, and the same margins funding Chevron’s dividends are feeding the price gauges that keep the Fed’s hawks dissenting.
Market Losers 📉
Meta (META). Shares fell 9.64% after earnings despite a revenue beat, because EPS missed by nearly 14% and free cash flow collapsed from $8.5 billion a year ago to $784 million as the company guided full-year capital spending to $130 to $145 billion. Meta became the first hyperscaler repriced this cycle specifically for spending without third-party customers attached. Investors should care because the market just published its rubric, and every AI budget without a receipt now gets graded against it.
Apple (AAPL). The stock slipped 3% to 4% after Thursday’s report, with one wire service quoting a 7% decline by Friday’s close. Services revenue of $30.74 billion missed the $31.22 billion estimate and Greater China came up short. Apple’s premium multiple rests on Services carrying the margin story while hardware carries the brand, so both engines sputtering in the same quarter raises a fair question about what supports the valuation while the company stays largely outside the AI capex race.
Coinbase (COIN). Coinbase reported a $359.5 million net loss, or $1.36 per share, on revenue of $1.2 billion against $1.35 billion expected, its third straight miss. Shares dropped about 6.5% in after-hours trading Thursday and roughly 10% on Friday per The Motley Fool. The company grew its share of industrywide trading volume to a record 10.3% while those volumes shrank more than 20%, which means it’s winning a bigger slice of a smaller pie. Investors should care because the stock functions as a leveraged bet on trading activity returning, and this quarter shows the cost of holding that bet while you wait.
What to Watch Next Week 👀
The July Jobs Report (Friday, August 7, 8:30 a.m. ET). The Fed’s 9-3 vote made this release the referee. Three dissenters want a hike, inflation has run above target for five years, and Q2 GDP growth just slowed to 1.5% from 2.1%. A hot payrolls number hands the hawks their argument for September; a weak one revives the case that the economy is already slowing under 3.5%-plus rates. Whatever prints Friday morning becomes the Fed’s operative fact.
Palantir (Monday) and AMD (Tuesday) earnings. The receipts audit moves downstream to AI software and chips. Palantir has guided quarterly revenue to $1.797 to $1.801 billion, and AMD’s data-center segment will show whether the hyperscalers’ hundreds of billions in spending is actually landing as chip revenue. Eli Lilly (Wednesday) and Disney round out the week. After Meta’s repricing, any AI-adjacent company reporting without demand evidence walks into a harsher room.
The Clarity Act’s Senate window. ZebPay notes the window for Senate passage of the crypto market-structure bill is narrowing, and this week showed the stakes concretely: the bill would clarify which tokens are securities and which are commodities, the exact ambiguity that leaves token holders outside the bankruptcy payment line described in this week’s article. Coinbase’s regulatory footing and every US token project’s legal status ride on whether it moves before the calendar closes.
Matthew Snider is the founder of Block3 Strategy Group, author of “Warren Buffett in a Web3 World,” and publisher of the BitFinance newsletter. He holds a Series 65 and MBA, and has been an active participant in digital asset markets since 2015. This article is for educational purposes only and should not be considered financial advice. Always consult with a qualified professional before making investment decisions.
Sources
block/buzz repository, README and architecture docs, Apache 2.0 — github.com/block/buzz
TechCrunch, July 21, 2026 — Buzz as a Slack and GitHub challenger
buzz-cli README, block/buzz — command surface and authentication
CNBC via Yahoo Finance: S&P 500 closes higher Friday as Amazon surges; Dow posts fourth straight winning month
Federal Reserve: FOMC statement, July 29, 2026
TradingKey: Microsoft +8%, Amazon +10%, Apple -4%: The Big Tech Earnings
Fortune: Chevron posts largest quarterly profit ever, Exxon income surges
Yahoo Finance: Coinbase Q2 2026 earnings miss: $359 million net loss
CoinDesk: Coinbase sinks 5% after missing Q2 revenue estimates
Coingabbar: How Bitcoin, Ethereum and XRP Performed in July 2026