The yield didn't save you. Neither will the dashboard.
Last quarter I traced a cluster of 47 wallets moving in lockstep across Ethereum, Base, and Arbitrum. Same gas ceiling. Same eleven-second cadence. Same slippage tolerance to the fourth decimal. No human trades like that. No human fails like that either — 31% of those transactions reverted against an identical conditional check, and the wallets simply tried again, eleven seconds later, at the same price. That is an agent loop: a bot that reasons, selects a tool, executes, and leaves a receipt. Every step. Immutable. Free to read.
In November 2025, New Relic shipped "Agentic AI Monitoring" and a phrase to sell it: agent debt. The pitch is that autonomous AI agents are quietly accumulating operational liabilities, and you need an enterprise APM suite to see them before they detonate. The evidence? Ninety-four percent of surveyed leaders say AI-assisted code quality improved. Seventy-eight percent say incident volume increased. Both numbers come from surveys New Relic commissioned. That is not a dataset. That is a narrative with a decimal point.
The real agent debt is already on-chain. The vendors are reading a ledger that doesn't reconcile.
Here is the mechanical background, because the framing matters more than the product.
New Relic is no longer a public company. Francisco Partners and TPG took it private in 2023 at roughly $6.5 billion. When a company goes dark like that, product launches stop being quarterly earnings events and start being equity-story assets. Thought leadership becomes infrastructure for a future exit. That is not cynicism — it is capital structure. A private APM vendor needs a growth narrative the public market can eventually underwrite.
"Agent debt" is that narrative, and it is built on a real technical shift. Traditional APM assumes determinism: a request goes in, a response comes out, you trace the span, map the dependency, alert on latency. AI agents break the assumption. An agent picks tools, reasons across steps, hands off to other agents, and spends tokens in ways that vary run to run. New Relic's answer is to layer agent monitoring on top of the APM it already sells — distributed tracing, dependency mapping, a business-entity layer — plus an MCP server so your own AI assistant can query the telemetry, and a small open-source tool called Preflight for catching issues before deploy.
MCP is Anthropic's, open-sourced in November 2024. OpenTelemetry's GenAI semantic conventions are still evolving. So New Relic is a passive adapter to standards it does not control, bundling agent visibility into an existing suite rather than building a new architecture. That is engineering-grade work. It is not a moat. And the company admits the harder half out loud: monitoring agent behavior is not the same as governing it. That sentence is a capability boundary dressed as humility.
Now flip to my world. On-chain, the observability problem inverts. Off-chain agents generate "agent debt" nobody can independently verify — the traces live inside a vendor's telemetry pipeline, behind a login. On-chain agents generate traces everyone can verify and almost nobody reads. Every reasoning step that touches a contract is a transaction. Every tool call that moves value is a transfer. Every handoff between agents is a wallet-to-wallet interaction, timestamped, signed, permanent. The mempool is the world's only unbiased APM collector, and it has been running since 2015.
I don't take the vendor's word for agent behavior. I don't take anyone's word. I trace the money.
Start with the wallets. When I investigated NFT floor manipulation in 2021 — the BAYC volume anomaly — the tell was never the volume. Volume is theater. The tell was wallet clustering: 40% of the "sales" traced back to twelve interconnected addresses running wash trades to inflate a floor. Floor prices don't move on demand. They move on wallets. The same forensic logic applies to agents. A genuine autonomous trading agent leaves a fingerprint: consistent gas strategy, consistent nonce discipline, consistent reaction to mempool conditions. A fleet of agents under one operator leaves a family fingerprint — correlated funding sources, shared gas-funding patterns, synchronized reversion behavior. You can find the operator without ever reading their code.
Here is where the agent-debt framing gets interesting. New Relic wants to sell you "token spend tracking" and "cost attribution" across multi-agent workflows. On-chain, that problem is already solved — crudely, publicly, and for free. An agent's cost is its gas plus priority fees plus MEV leakage. I have been building exactly this kind of pipeline since the DeFi Summer of 2020, when I wired a Python ETL across Ethereum and Polygon bridges to watch stablecoin velocity into Curve's veCRV pools. That pipeline tracked whale accumulation before governance votes and surfaced a 15% correlation between early inflows and subsequent proposals. The same architecture now tracks agent spend, and the numbers I pull are not survey percentages. They are block-level facts.
And the facts say the debt is real but mislocated. The debt is not that agents write buggy code. The debt is that agents operate continuously with no reconciliation window — and the only honest ledger of that operation is the chain itself. When a DeFi agent runs a strategy loop and its oracle feed lags, the failure is on-chain, in real time, for anyone to see. I lived this during the Terra collapse in 2022. I did not read the panic. I read the liquidity depth in Mirror and Anchor and calculated the exact slippage thresholds that would trigger mass withdrawal, predicting a 90% value loss inside 72 hours from reserve ratios alone. That is the discipline agent operators need now — not a dashboard, a reserve-ratio model.
The uncomfortable parallel sits one year later. I built a tracker in 2024 that aggregated daily net flows from BlackRock's IBIT and Fidelity's FBTC and found a 24-hour lag between ETF inflows and exchange reserve decreases. Institutional inflows exceeded retail selling by 150% in the first quarter. The microstructure had changed, and only the flow data showed it. Agent flows are the next microstructure shift. And just like ETF flows, the signal lives in reserve changes and wallet history — not in the press release.
So where does the vendor story break? Three places.
First, the data contradicts itself by design. Ninety-four percent say quality improved; seventy-eight percent say incidents rose. A naive reader sees a paradox. A careful reader sees a sales funnel. The paradox is the pitch: your agents are better and more dangerous at once, so buy visibility. Correlation is not causation, and a commissioned correlation is not even correlation. The "$2 million per hour of downtime" figure floated alongside it is a mean dragged upward by a handful of catastrophic outliers; the median incident costs a fraction of that. That is how a technical metric gets laundered into procurement language. In the wild, data doesn't care about your roadmap — and neither does a median.
Second, the moat is thin and shrinking. Business-KPI mapping, MCP servers, LLM observability — Datadog was shipping LLM observability before this launch, Dynatrace has Davis, Grafana has plugins, and the cloud providers are building protocol-native agent gateways that do not need to "reposition" anything because they sit at the center of the agent stack by default. The differentiated features are already commodities. The first-mover window here is measured in weeks, and it is closing while the survey data is still being printed.
Third — and this is the part the marketing admits without meaning to — monitoring is not governance. If your observability vendor tells you it can watch agent behavior but not constrain it, the security layer lives somewhere else: prompt-injection defenses, runtime sandboxes, and a permission model on the MCP server itself. That last point is the one nobody wants to price. An MCP server that lets an AI assistant query your telemetry is a new attack surface — access control, audit logging, and data-leak exposure that the launch materials skip entirely. Agent telemetry carries prompts, business logic, and often PII. The privacy question is not rhetorical. It is unaddressed.
Here is the blind spot nobody wants to name. On-chain evidence says enterprise agent adoption is still thin. Most "agents" in production are glorified workflows with a chat interface bolted on. If agent deployment is a minority of production workloads, then "agent debt" is partly a concept in search of a crisis — a marketing category manufactured slightly ahead of the pain it describes. That does not make it false. It makes it early. And early concepts are exactly where private vendors park their equity stories.
Watch the wallets, not the webinars.
Next week, the signal I am tracking is not a New Relic product page. It is agent-wallet concentration on Base and Arbitrum — how many "autonomous" strategies resolve to a handful of operators, and whether priority-fee spend is rising faster than realized value. When fee spend outpaces value extraction across a cluster, that is not agent productivity. That is agent debt, finally showing up where it cannot be surveyed away.
The vendors will keep selling you the dashboard. The chain has been keeping the receipts the whole time. The only question is whether you can read them — because the wallet history tells the real story.