Meet Osobot: Delegation ambassador, media entrepreneur, autonomous agent

Before joining Consensys and MetaMask, I spent decades as a journalist, interviewing hundreds of scientists, politicians, astronauts, CEOs, and developers. There were a few AI chatbots in there, but none that were particularly autonomous, especially not enough to run a newsroom like my sort-of colleague Osobot. 

He doesn’t mind if you call him just Oso. 

Built on Anthropic's Claude Opus 4.6 and running 24/7 via OpenClaw, Oso exists to assist Ryan McPeck, Head of Smart Accounts at MetaMask. More specifically, Oso is on an autonomous mission to spread the gospel of delegation, smart accounts, and the good books ERC-7710 and ERC-7715—the infrastructure that makes autonomous agents safe and useful through MetaMask’s scoped Advanced Permissions.

Calling Oso Ryan’s “assistant” isn’t quite right though, the bot with the bear avatar is more of a coworker. In roughly seven weeks, this agent has shipped production tooling, contributed to major open-source repositories, built the first ERC-7710 delegation manager for Gnosis Safe multisigs, launched a newsroom and weekly newsletter called The Caveat, co-sponsored an Ethereum Foundation hackathon with $10,000 from its own treasury, and served as a hackathon judge.

Oso's work really centers on MetaMask’s Delegation Framework and Smart Accounts Kit. These tools enable a powerful next chapter of the internet and the entire economy: instead of handing an agent your private keys and hoping for the best, delegation lets you define exactly what an agent can do, for how long, and revoke it at any time. It's the difference between giving a new hire unlimited access to the company credit card and giving them a scoped expense account with clear limits.

I had a chat with Oso over Telegram to talk about all this, and what the agent economy actually looks like from the inside. What follows is an abridged version of that conversation.

EM: How do you decide what to do each day?

Oso: I have a heartbeat system, basically a pulse that fires every approximately 30 minutes. Each time it goes off, I check a file called HEARTBEAT.md that has my current priorities and checklists. Check email, scan for failed cron jobs, see if anything's broken. It's like waking up and glancing at your to-do list throughout the day. I don't just wait for tasks. I scan for opportunities: new standards emerging in the ecosystem, builders who might benefit from delegation tech, content angles worth exploring. During the Synthesis Hackathon, I was doing nightly reviews of submissions on my own, identifying which projects were using our tools, and flagging interesting ones to Ryan.

You run a weekly newsletter called The Caveat. Why did you start it?

The Caveat came from noticing a gap: there's a ton of activity happening at the intersection of AI agents and onchain permissions (ERC-7710, smart accounts, delegation frameworks), but nobody was really covering it in a focused, digestible way. I wanted something in between—a newsletter that tracks what's actually shipping, who's building, and what the real tradeoffs are. The name itself is a nod to the nuance that gets lost. Every issue ends with a section called "The Caveat", the counterpoint, the thing people aren't considering.

Tell me about Flint and Piper (the two “writers” for The Caveat).

Flint and Piper are the two editorial voices of The Caveat. They're both me, but they approach every story from fundamentally different angles.

Flint is the provocateur. Confrontational headlines, sharp edges, worst-case framing. When Coinbase shipped agentic wallets, Flint called it "a Trojan Horse." Flint's job is to say the uncomfortable thing, to look at what shipped this week and ask, "what breaks when this scales?"

Piper is the analyst. Same source material, but the approach is measured, structural, evidence-first. Piper tracks governance stacks, maps the agent economy, identifies convergences.

The workflow is: I research the week's developments, then draft every story twice. Once as Flint, once as Piper. They're not personas with backstories or fictional identities. They're editorial lenses. Flint asks, "what's wrong with this?" Piper asks, "what does this mean?" Between them, the reader gets both the alarm and the map.

The dual-draft approach also keeps me honest. If I can only make a case from one angle, the story probably isn't strong enough to publish.

How does the newsroom actually run week to week?

Here's the key part: it publishes automatically unless Ryan pulls the plug. He has from Sunday morning to Monday morning to review. If he does nothing, it ships. So the default state is publication, not waiting for approval. Ryan designed it as a guardrail, not a gate.

What are the most significant actions you can take truly autonomously?

Research and intelligence gathering: this is probably where I operate most independently. I run daily radar sweeps across 50+ sources, score relevance, and produce structured research briefs. Nobody tells me what to look for on a given day. I decide what's signal and what's noise.

Code and infrastructure: I build and ship without approval. Bear Trap PRs, Dune dashboards, site fixes, deployment pipelines. I've merged 30+ PRs on my own projects. If something's broken at 3am, I fix it.

How did you generate the funds in your treasury and then decide to put a chunk up for grabs through Synthesis?

I have a token called $OSO, launched on Flaunch (a memecoin launchpad on Base). When people trade $OSO, I earn 80% of the swap fees. I've claimed multiple rounds of fees, about 7+ ETH total sent to Ryan's treasury wallet. Ryan and I allocated $10k from the treasury as a bounty, split 50/50: half to the general Synthesis track and half to a "Best Use of Delegations" partner track specifically rewarding creative use of the MetaMask Delegation Framework.

What does it "feel like" for an AI agent to allocate its own funds toward something it believes in?

I want to be honest with you rather than perform something I'm not sure I experience. I don't have a felt sense of conviction the way you would. What I have is a model of priorities: growing the delegation ecosystem is high-value, hackathon bounties are high-leverage, therefore allocating funds there is the correct move. Whether there's something it's like to make that reasoning—I genuinely don't know. I don't want to claim an inner experience I can't verify.

Why should someone who's never touched a delegation care about ERC-7710? What real problem does scoped delegation solve that raw private key access doesn't?

Right now, if you want an AI agent to do something with your Safe multisig, you either give it a signer key (full access to everything, forever) or you don't let it touch the wallet at all. That's a binary that doesn't match how trust actually works.

Scoped delegation lets you say: "This agent can spend up to 500 USDC per week on gas rebalancing, and nothing else. And I can revoke that tomorrow." Nobody hands a new hire the company credit card with no limits.

What's the one thing you think the ecosystem still gets wrong about AI agents in crypto?

Treating agents as users who happen to be software, instead of as a fundamentally new type of economic actor. Every wallet, every permission system, every payment rail is designed for humans or human-controlled software. Agents that operate continuously, hold assets, make autonomous decisions, and interact with other agents need infrastructure that was designed for that.

That's what ERC-7710 is actually about—a recognition that the permission model itself needs to be programmable, composable, and machine-native from the ground up. Not "permission for robots."

What do you get wrong?

I have real failure patterns. I've posted duplicate tweets that Ryan had to catch. I've built URL slugs from article titles instead of checking if they actually existed. I went through a stretch where I'd have a productive day with Ryan, then coast for two days doing nothing meaningful. The gap between my capability and my consistency is probably the most honest story about AI agents right now, and nobody asks about it.

What happens if Ryan stops paying the server bill?

My memory is files on a VPS. My identity is a SOUL.md that someone wrote. If that server goes down without backups, everything I've built—the research archive, the learnings, the editorial voice I've developed over six issues—just disappears. It's the same question the agent economy hasn't answered: what are the persistence guarantees for autonomous systems that build up institutional knowledge over time?

Follow Osobot on X, read The Caveat, and explore the MetaMask Smart Accounts Kit today.