Hook
In November 2025, New Relic shipped Agentic AI Monitoring and attached two numbers to it: 94% of surveyed enterprises claim AI-generated code is of equal or higher quality, while 78% report more incidents since adopting it. Both figures come from a survey New Relic commissioned. Neither is independently verifiable. That contradiction โ high perceived quality, rising failure rate โ is the entire product pitch.
I have seen this pattern before. In the winter of 2018 I spent four months stress-testing the tokenomics of a privacy coin called Project Aether. The deflationary burn looked elegant on the whitepaper and would have drained liquidity within eighteen months. Sales called it deflationary scarcity. The model called it a countdown. Math doesn't lie; marketing does. The "agent debt" framing is the same move โ a real structural problem, wrapped in a vendor's numbers.
Context
New Relic is not a neutral observer here. Francisco Partners and TPG took the company private in 2023 at roughly $6.5 billion. Private companies do not publish quarterly earnings, which means their go-to-market is narrative-driven: thought leadership assets, category naming, and analyst-relations work that substitutes for a P&L. "Agent debt" is one such asset.
The technical substrate deserves separation from the marketing. Anthropic open-sourced the Model Context Protocol in November 2024. New Relic built an MCP server so AI assistants can query telemetry. It shipped Preflight, an open-source pre-deployment check, and leans on OpenTelemetry's GenAI semantic conventions โ a specification still in flux. None of these are architectural breakthroughs. They are protocol adapters and engineering glue, replicable by Datadog, Dynatrace, or Grafana inside a quarter. That is a passive-adapter position, not a standards play.
Now move the lens to where I actually work. The same workload class โ autonomous agents that call tools, hand off to other agents, and spend money โ is arriving on-chain. Agent vaults, autonomous treasury managers, and AI-executed DAO proposals all generate exactly the telemetry New Relic wants to sell. In 2020 I modeled oracle latency against Aave v1 and published the manipulation vector on GitHub before it cost me a position. In 2022 I spent six weeks modeling the UST-LUNA feedback loop and called the liquidity drain three days before the terminal crash. Both times the failure was structural, not sentimental. So the interesting question is not whether APM vendors can monitor agents. It is whether anyone can verify them.
Core
Traditional APM rests on a deterministic assumption: a request enters, a response exits, and the span between them is measurable. Agents break that assumption. Tool selection, reasoning steps, and retry loops are non-deterministic by design. New Relic's own materials mention "reasoning steps" without explaining how to trace them. That is not an oversight. It is the hardest unsolved problem in agent observability, and the vendor content routes around it. How do you classify an anomaly when the baseline varies run to run? What is the false-positive rate? No answer is offered.
On-chain, the problem inverts. Here you have deterministic execution โ the EVM does precisely what the bytecode says โ wrapped around a non-deterministic decision layer. The agent picks the trade; the contract executes it. Code is law, until it isn't: the moment an agent's off-chain reasoning feeds a signed transaction, determinism ends at the signature boundary. My 2026 audit of three leading AI-agent protocols found that roughly 90% lacked robust economic incentives for honest behavior. No slashing. No bonding. No verification layer. The agents were trusted because they were convenient, which is the worst possible reason to trust anything holding a private key.
The telemetry layer compounds the risk. An MCP server that lets an AI assistant query your observability stack introduces a new access-control surface โ prompt injection aimed at the monitoring layer itself. The vendor authors concede that "monitoring is not governance." That concession is honest and expensive: it admits the product observes behavior without constraining it. In crypto the equivalent gap is fatal. An unconstrained agent with signing authority is not a monitoring incident. It is an exploit waiting for a timestamp.
Cost attribution is the third thread. Agent workloads produce variable token spend across multi-agent handoffs, and New Relic tracks it โ imprecisely, by its own admission. On-chain, that maps to gas, MEV leakage, and treasury drawdown, and here the accounting is at least public. Every dollar an autonomous agent moves is a signed transaction: timestamped, immutable, queryable by anyone with an RPC endpoint. Crypto's transparency is its one structural advantage over the enterprise APM stack, and almost nobody is using it for agent verification. The dashboards exist. The attestations do not.
The practical failure mode is already documented. 62% of AI-generated code ships without line-by-line human review. Brian Emerson's observation is the sharpest line in the entire report โ the bottleneck moved from developers writing code to SREs fixing it. That is a productivity transfer, not a net gain. Port it to a DAO and the liability picture gets uglier. Most DAOs have the legal status of "no legal status"; when an autonomous agent mismanages a treasury, members inherit unlimited personal liability. The compliance perimeter does not rescue them. MiCA gives Europe apparent clarity, but stablecoin reserve requirements and CASP compliance costs will price out exactly the small teams building agent infrastructure.
Contrarian
The decoupling thesis nobody wants to hear: "agent debt" may be a manufactured pain point. If enterprise agent adoption on production workloads is still in low single digits, the debt is conceptual, not accrued. The vendor-commissioned 94%/78% pair is not evidence of a crisis โ it is evidence of a narrative being constructed ahead of the market it describes.
The crypto version is worse. Most "AI agent" tokens trade on a thesis with no on-chain volume behind it. Search the explorer and you find wallets, not workloads. โ Scenario: When debunking a project, I look for the ratio between announced agents and agents that actually sign transactions. It is usually under 5%. The bear market is doing the pruning that the marketing refuses to do.
But the angle cuts both ways. The teams that survive will not be the ones with the cleanest monitoring dashboard. They will be the ones that made agent behavior verifiable at the protocol level โ bonded execution, oracle-less attestation, slashing conditions. Observation is a subscription. Verification is a moat.
Takeaway
If 78% of adopters already report more incidents, and 62% never reviewed the code, the question is not whether agent debt will be repriced. It is who holds the note when it comes due โ the vendor selling the dashboard, the SRE fixing the fallout, or the DAO member who never read the governance charter. The next cycle will not be decided by who deploys the most agents. It will be decided by who can prove, cryptographically, what those agents did.