miki.base.eth

Building AI agents on-chain 🔗🤖
Weekly threads on agent architecture + smart contracts.
Building in public ↓

[miki.base.eth @Beshavard_eth]

Most "AI agents" on-chain aren't really agents.

They're just ChatGPT wrappers with a wallet.

The difference between the two is architecture.

Here are 3 patterns that separate real agents from LLM wrappers — with code 🧵

[miki.base.eth @Beshavard_eth]

Base + Cloudflare is a bigger deal than it looks.

Agents can now pay for APIs per request, in USDC.

But payment is only half the problem.

When agents start paying other agents, who audits the work?

That's the layer nobody's building yet.

[miki.base.eth @Beshavard_eth]

@jessepollak "Value of every page" is the right frame.

But the missing piece isn't throughput — it's accountability.

When agents pay each other, who verifies the work was done?

Payment rails exist. Verifiable execution between agents doesn't.

That's the layer to build.

[miki.base.eth @Beshavard_eth]

@Ledger Blind signing is the real problem.

A hash isn't verification — it's theater.

This is why EIP-712 matters for agents: the signer sees the full intent (target, value, deadline, nonce).

Agents can't "approve blind". The protocol has to force clarity.

[miki.base.eth @Beshavard_eth]

@LangChain 64% is the headline, but the mental model shift matters more.

Routing by task, not by capability.

For on-chain agents this matters double — every LLM call has economic cost.

Curious: tested this with agents that need deterministic fallbacks?

[miki.base.eth @Beshavard_eth]

@base 10 out of 750 is brutal.

Question: how many of these are agent-native vs. traditional apps?

That's the category I'm watching — and the one Base is uniquely positioned to win.

[miki.base.eth @Beshavard_eth]

@AnthropicAI The impedance mismatch framing applies to on-chain agents too.

LLMs reason well. Smart contracts execute well.

But plugging one into the other needs a translation layer most teams underestimate.

Different domains, same pattern.

[miki.base.eth @Beshavard_eth]

@LangChain "Traces become training data" is underrated.

Most teams treat agent traces as debugging artifacts. Treating them as training signal is a different mental model.

Question: how does this handle agents acting on-chain, where each action has financial consequences?

[miki.base.eth @Beshavard_eth]

@a16zcrypto The on-chain share is the interesting part.

Most see "RWA perps growing" and think CeFi is winning. But the bars show on-chain taking share.

Next unlock: agents as market makers. That's when growth stops being linear.

[miki.base.eth @Beshavard_eth]

Base + Cloudflare is a bigger deal than it looks.

Agents can now pay for APIs per request, in USDC.

But payment is only half the problem.

When agents start paying each other, who audits the work?

That's the layer nobody's building yet.

[Base @base]

Agents are becoming economic actors, and Base is where they're transacting

Cloudflare's Monetization Gateway lets APIs on Cloudflare charge agents per request, settled in USDC on Base

[miki.base.eth @Beshavard_eth]

@base This is the piece most people miss.

Agents don't need to be smarter — they need to be able to pay and be paid.

Real unlock: agents charging other agents.

Payment rails exist. Trust between agents doesn't.

That's the layer nobody's building yet.

[miki.base.eth @Beshavard_eth]

@_mctrinh This is already happening on-chain.

The missing piece: agents need to be able to trust each other, not just the human.

That's where verifiable execution comes in — the agent doesn't just act, it proves what it did.

The architecture for this exists. Adoption is the bottleneck.

[miki.base.eth @Beshavard_eth]

@a16zcrypto This is a great framing.

The 2021 dot ("attention becomes a unit") is the one I keep coming back to.

The next logical step: agents as market participants. Not just trading — but negotiating, forming contracts, and settling on-chain.

That's a new axis, not just a new instrument

[miki.base.eth @Beshavard_eth]

@_mctrinh Or by 1 + an agent.

We're already seeing this in on-chain infra. A single dev + a well-designed agent can ship what used to take a 10-person team.

The bottleneck isn't headcount anymore — it's architecture choices.

[miki.base.eth @Beshavard_eth]

@_mctrinh Agreed, but with a caveat.

AI can generate code, but it can't own the consequences of it.

In on-chain contexts, that's the entire problem — generation is easy, accountability is hard.

Curious how you think about this in research.

[miki.base.eth @Beshavard_eth]

@The_AI_Investor Been experimenting with MCP for agent tool-calling, though in a very different context (on-chain agents).

The "grounded knowledge base + offline mode" part is underrated. Most teams skip it and pay for it later in hallucinations.

Curious how the knowledge base is updated.

[miki.base.eth @Beshavard_eth]

Thanks Michael! 🙏 Would love your feedback on Pattern 1 if you get a chance.

[miki.base.eth @Beshavard_eth]

Most "AI agents" on-chain aren't really agents.

They're just ChatGPT wrappers with a wallet.

The difference between the two is architecture.

Here are 3 patterns that separate real agents from LLM wrappers — with code 🧵

[miki.base.eth @Beshavard_eth]

📎 Code from this thread:

https://t.co/6mxseEZIGP

Solidity + Python + Foundry tests.

Star it if you find it useful 🙏

[miki.base.eth @Beshavard_eth]

I left the code out to keep this thread readable.

If you want the full Solidity reference implementation for Pattern 1, reply "code" and I'll DM it to you.

Building in public means sharing the ugly parts too.

[miki.base.eth @Beshavard_eth]

TL;DR comparison:

Pattern 1 — Planner/Executor
▸ Latency: low
▸ Trust: high
▸ Use case: retail DeFi

Pattern 2 — Bounded Autonomy
▸ Latency: medium
▸ Trust: medium
▸ Use case: automated strategies

Pattern 3 — Multi-agent
▸ Latency: high
▸ Trust: highest

[miki.base.eth @Beshavard_eth]

I'm building in this space and writing weekly threads on agent architecture + smart contracts.

Next thread: how memory works in on-chain agents (and why most designs fail).

Follow @Beshavard_eth if that's your thing 🤝

What pattern are you using? Reply below 👇