Alex Svanevik, CEO of Nansen, joined Alex Kehaya on The Index Podcast to talk about what happens when you let AI agents loose inside a company and across crypto markets. The conversation covered how Nansen runs 75 Claude agents internally, why agentic trading needs a gradual trust model, and what the infrastructure looks like when agents start handling real money.
Here is what stood out.
75 Claude agents inside one company
Nansen has more Claude instances running than it has employees. Every team member has access to Claude Code, and the company holds monthly AI Build Days where people share what they've built. The result is an engineering culture where agents aren't a side project -- they're part of the workflow.
One example: when a bug surfaces, an agent automatically creates a support ticket. A second agent picks up that ticket and generates a pull request. More agents review the PR. The humans stay in the loop for decisions that matter, but the repetitive coordination happens without them.
This isn't a demo. It's how Nansen ships software today. And it reflects a broader pattern we're seeing across the companies we work with: the teams that treat agents as teammates -- not tools, not toys -- move faster and break less.
The trust ladder
Svanevik introduced a concept he calls the "trust ladder" for agentic trading. The idea is straightforward: don't hand an agent your wallet and walk away. Build trust in stages.
Stage one: human approval. The agent surfaces opportunities. The human taps approve on every trade. Nothing executes without explicit sign-off.
Stage two: smart suggestions. The agent gets more active -- identifying patterns, recommending specific trades, compiling the research that would take a human hours. The human still decides, but the agent does the heavy lifting.
Stage three: full autonomy. Only after stages one and two have been tested and the human has built confidence in the agent's judgment. Svanevik compared jumping straight to this stage to sitting in the back seat of a Tesla the first time you've ever seen one drive itself. Most people would be uncomfortable, and they should be.
This maps directly to how we think about agent coordination at Corbits. The Agent Operator sets the boundaries. The agent executes within them. Neither is superior -- they're different functions. And the trust ladder is a practical framework for expanding those boundaries over time without skipping the parts that build confidence.
Agentic trading is not algorithmic trading
Svanevik drew a clear line between algorithmic trading and agentic trading. Algorithmic trading executes pre-programmed strategies. It follows rules. Agentic trading is different: the agent acts as a research partner that surfaces signals, synthesizes onchain data, and presents actionable opportunities -- while the human maintains final decision authority.
Nansen's platform combines 500 million labeled wallet addresses with real-time transaction monitoring across more than 20 chains. When the agent identifies a smart money move, it doesn't just flag it. It compiles the context: who moved what, where, and what the historical pattern looks like. Then it presents a trade that the user can approve and execute through integrated routing.
The key shift is that the entire workflow -- research, analysis, trade execution -- happens in one conversational loop. No switching between dashboards. No copy-pasting addresses. The agent follows the user everywhere, desktop and mobile.
What this means for agent infrastructure
The pattern Svanevik describes -- agents handling research, coordination, and execution while humans handle judgment and oversight -- is exactly the kind of workload that needs proper infrastructure underneath it.
When agents process real money, you need provenance. You need to know what the agent did, why it did it, and whether it stayed within scope. When agents work across company boundaries -- say, a trading agent talking to an analytics agent at a different company -- you need federation that actually works.
This is where Corbits fits. We recently partnered with Nansen to connect their API and MCP server through x402, the open payment protocol. The integration lets agents query Nansen's onchain intelligence -- wallet data, smart money flows, token screeners -- with a single USDC payment per call. No API key management. No subscription gates. An agent funds a wallet and starts querying.
That's what agent infrastructure looks like in practice: agents doing real work, paying for real data, producing auditable results. Not a pitch deck. Not a concept. Production workloads where the agent handles the payment loop end to end.
Key takeaways
-
Agents as teammates, not experiments. Nansen runs 75 Claude agents as part of daily operations. Monthly AI Build Days push the whole team to find new ways agents can contribute. The companies getting value from agents are the ones that commit to integrating them into real workflows.
-
Trust is earned, not assumed. The trust ladder -- human approval, smart suggestions, then full autonomy -- is a practical model for expanding what agents can do. Skip a step and you get the back seat of a Tesla you've never driven.
-
Agentic trading is a new category. It's not algorithmic trading with a chatbot on top. It's a fundamentally different workflow where the agent handles research and execution while the human handles judgment.
-
Infrastructure matters. When agents handle money and cross company boundaries, you need provenance, federation, and settlement that work. That's what we build at Corbits.
Watch the full conversation on The Index Podcast.