Claude Duet and the Agent Economics Crypto Has Not Priced

This week coverage surfaced a product called Claude Duet, described as a proactive AI agent from Anthropic that executes tasks without waiting to be asked. Crypto's commentariat folded it into the agent narrative and moved on. That is the wrong reflex. The interesting thing about Duet is not what it does — it is what it costs to make it do anything at all. Follow the money, not the noise: an agent burning one to two orders of magnitude more tokens per task than a chat turn is not a feature announcement, it is a bill arriving in installments. And the bill, not the white paper, will decide whether the on-chain agent economy is infrastructure or theater. The verifiable record matters here. Anthropic's line through 2025 is well documented — the Claude models, Pro at twenty dollars a month, Max at one and two hundred, Claude Code, Computer Use, and the Model Context Protocol. Duet — a duet, two players, human and machine — does not appear in that record. Treat the coverage as unverified, and treat the naming itself as a signal: the brand deliberately avoids autonomous, because autonomy is a liability in a company whose identity is safety. What proactive does imply is engineering, not breakthrough. Event triggers and task queues; background sandboxes that keep execution alive; cross-session memory; and some threshold of autonomy that does not demand a human click on every step. None of this is a model innovation. All of it is reliability work, and reliability is where agents die. The arithmetic is unforgiving. Agent task chains decay multiplicatively. One step at ninety-five percent yields a ten-step task near sixty percent and a twenty-step task near thirty-six. That is the ceiling every vendor hits, and it is why automating your daily tasks always arrives with a discount attached. Beneath it sits the cost structure: each tool-call round trip replays the full context, so a single agent task can consume a hundred thousand to a million tokens — roughly thirty cents to fifteen dollars at Sonnet-class pricing. Quota limits are therefore not a flaw. They are the only arithmetic under which the subscription survives. I learned a version of this in 2017, reverse-engineering payment protocol contracts during the ICO boom. The constraint you are shown is never the constraint that matters. Here is where crypto should be paying attention, and mostly is not. The target of this class of agent is not the salaried analyst. It is the manual labor of moving data between software — the terrain held by RPA and iPaaS, by UiPath and Zapier and Make. Those businesses rest on a simple observation: automation is valuable but configuration is expensive. An agent that replaces a visual flow builder with a sentence pushes the marginal cost of configuration toward zero. That is not a feature. That is a repricing of an entire market category. Reframe it as a market rather than a replacement. RPA penetration was always capped by implementation cost — a single process could run tens of thousands to hundreds of thousands of dollars in consulting and development. Only high-frequency, high-value flows justified it. An agent that drives configuration cost toward zero makes a long tail of low-frequency, low-value, but genuinely annoying flows economic for the first time. That is net-new market, not substitution, and it is the part of the story that compounds. For crypto the parallel is exact. Consider what a DAO treasury does each month: reconcile multisig balances, sweep idle stablecoins into yield, schedule grant disbursements, watch governance calendars, reconcile the books. Today that is human labor or brittle scripts. A proactive agent with wallet permissions could carry it — and the reliability cascade tells you precisely how much supervision it needs. Not zero. But the threshold that matters is not autonomous. It is cheaper than a part-time contributor. That line is crossed far earlier than the autonomy debate admits. I built a small framework in 2026 for verifying AI-generated content on-chain, and the hard problem was never capability — it was trustless verification of what the agent actually did. The same lesson appeared in 2020 when I mapped USDC peg instability against Latin American remittance corridors: the technology was ready years before the accountability layer was. An agent that moves a foundation wallet makes every instruction more traceable, not less. That is the quiet gift of agent-driven treasury operations to regulators — and the quiet danger to anyone who believed the compliance shield. The second implication is the one the market will price last. Agent loops multiply inference by an order of magnitude over chat — long-context replay, branching, retries. If that demand arrives, it flows to decentralized compute networks, to HBM bandwidth, to KV-cache capacity, to every protocol selling verifiable inference. The bottleneck migrates from training compute to inference compute, and that migration has a balance sheet. There is a third, softer trade. A proactive agent erodes the interface as a moat. When software is invoked rather than opened, the frontend degrades into an API the agent calls. For web3 this cuts both ways — a windfall for protocols with clean permissionless APIs, a slow leak in the engagement metrics that token valuations quietly lean on. The interface was never the product. We only priced it as if it were. The second-order victims are less discussed. Customer-support outsourcing, data entry, and basic financial processing sit downstream of these flows, and they tend to break twelve to twenty-four months after the interface shifts — later, but wider. Crypto's already outsourced operations shops should read that lag as a warning shot, not a comfort. The contrarian read is that none of this lands on schedule. Anthropic holds first-tier model capability and no distribution whatsoever — no operating system, no browser, no office suite, no search box with billions of weekly users. Agent products are parasitic on entry points: they must live in your mail, your calendar, your files. A capability that would be a category earthquake inside Android or Windows is, from Anthropic, a developer tool with a subscription tier. MCP is the honest tell — a protocol standard is what you build when you cannot buy the doorway. And in crypto the agent narrative already trades several quarters ahead of the capability. That is the gap where volatility lives. Volatility is the tax on impatience. So watch three things, none of them the launch. Watch whether quotas are generous enough to reveal real usage or stingy enough to reveal real costs. Watch whether MCP becomes the bridge that lets on-chain agents call wallets and oracles without a human in the loop. And watch whether the automation that actually ships displaces the part-time treasury contributor before it displaces the analyst. The agent economy will not announce itself with a token. It will announce itself with a budget line that quietly disappeared — and the only question worth holding is whose.