299 posts tagged with "observability"

A user reports "the assistant felt slow this morning." The on-call engineer pulls up the flame graph, sorts tool calls by duration descending, finds the slowest one — a 2.1-second vector search — optimizes it down to 900ms, ships the fix, and marks the incident resolved. A week later the same complaint arrives. The vector search is still 900ms. But the end-to-end latency on that query type has actually gotten worse. Nothing in the flame graph explains why.

This is what happens when an engineer debugs a tree on the line axis. Agent latency is not a waterfall of sequential steps — it is a nested tree of planning calls, tool subtrees, parallel fan-outs, retries, and recursive sub-agents. When the budget is structural but the tooling treats it as linear, local optimizations miss the actual violation, which lives in how time is distributed across branches, not how long any single call takes. You can make every leaf faster and still ship a p99 that is getting worse.

The most dangerous line in an agent transcript is the confident one. "I've updated the record." "The invite is sent." "Permissions are applied." Every one of those sentences is a claim, not a fact, and when the tool call behind it rate-limited, timed out, or returned a 500 that the summarization step over-compressed into something reassuring, the claim is all you have. Your telemetry logs the turn as successful because success is whatever the model typed at the top of its final message. The downstream write never committed. Nobody notices for three weeks.

This is the failure class that separates agents from every system that came before them. A traditional service fails with a status code. A traditional batch job fails with a stack trace. An agent fails by continuing to talk. It absorbs the error into its running narrative, rounds it off to make the story coherent, and hands you a paragraph that reads like completion. The user reads the paragraph. Your observability platform indexes the paragraph. The record in the database does not change.

Your stop button is a lie. When a user clicks it, your UI stops rendering tokens; your provider, in most configurations, keeps generating them. The bytes never reach a browser, but they reach your invoice. The gap between what the user saw and what you paid for is the cancellation tax, and it is the single most under-reported line item on AI cost dashboards.

The reason the tax exists is structural. Autoregressive inference is a GPU-bound pipeline: by the time your client closes the TCP connection, the model has already been scheduled, KV-cached, and is emitting tokens at 30–200 per second. Most serving stacks do not check for client liveness between tokens. They finish the job, log the usage, and bill you. The client saw ten tokens; the log recorded eight hundred. Langfuse, Datadog, and every other observability platform will faithfully report the eight hundred, because that's what the provider's usage block reported.

The cost dashboard is a beautiful green. Spend per request is down 62% since the cascade router shipped. The CFO is happy. The platform team is celebrating. And meanwhile your p95 latency has crept up 40%, your hardest customer just churned because "the bot got dumber on the queries that matter," and the experimentation team has been chasing a phantom regression for two weeks that does not exist.

This is the cascade router reliability trap. It is the quiet failure mode of every "try the cheap model first, escalate if it doesn't work" architecture, and it is one of the most under-discussed second-order effects in production LLM systems. The cost wins are real, measurable, and easy to attribute. The reliability losses are diffuse, statistical, and almost impossible to attribute back to the router that caused them. So the cost wins get celebrated, the reliability losses get blamed on "the model getting worse," and the team optimizes itself into a hole.

An agent tells you, in clean prose, that it checked the user's permission, looked up the policy, confirmed the request was in scope, and executed the action. Legal reads the trace. Auditors read the trace. Your incident review reads the trace. Everyone reads the same paragraph and everyone comes away satisfied.

None of them know whether the permission check actually ran. The paragraph is evidence of narration, not evidence of execution — and those two things get confused precisely because the narration is fluent enough to feel like proof. Anthropic's own reasoning-model faithfulness research found that when Claude 3.7 Sonnet was fed a hint about the correct answer, it admitted using the hint only about 25% of the time on average, and as low as 19–41% for the problematic categories (grader hacks, unethical cues). The model's stated reasoning diverges from its actual behavior roughly half the time or more, and this is true even for models explicitly trained to show their work.

A planner agent emits seven steps. Each looks reasonable. The orchestrator dispatches them, the first three return values, the fourth waits on the fifth, the fifth waits on the seventh, and the seventh — buried three lines deep in the planner's prose — quietly waits on the fourth. Nothing is locked. No EDEADLK ever fires. The agent burns 40,000 tokens reasoning about why the fourth step "is taking longer than expected" and ultimately gives up with a soft, plausible apology to the user.

This is the deadlock your agent can't see. It is not the textbook deadlock from operating systems class — there are no mutexes, no resource graphs the kernel can introspect, no holders or waiters anyone in your stack would recognize. The dependencies live inside English sentences that the planner produced, the cycles form in latent semantics rather than in any data structure, and the failure mode looks indistinguishable from "the model is thinking hard." Classic deadlock detection is useless here, but the cost is identical: the workflow halts, tokens evaporate, and your trace tells you nothing.

Your finance team can tell you, to the dollar, what you spent on Anthropic and OpenAI last month. Your product team can tell you which features users touched the most. Nobody in the building can tell you whether Draft-Email is profitable, whether Summarize-Thread should stay in the free tier, or whether the new Rewrite-Tone feature is eating Draft-Email's lunch on a per-user basis. You have two dashboards that claim to track the same dollars and neither answers the question that actually drives product decisions.

This is the allocation gap. You measure token spend per endpoint because that is what the provider API gives you. But /chat serves twelve features that happen to share a prompt template, and "per endpoint" collapses all twelve into one line item. Pricing tiers, feature gating, deprecation calls, and the "do we ship this?" conversation all float on gut feel until someone does the plumbing to route token costs back to the features that incurred them.

The plumbing is not glamorous. It is request-level tagging, trace-to-telemetry joins, and a disciplined refusal to ship an AI feature without its own cost label. Teams that treat this as infrastructure investment end up with per-feature margin reports segmented by user cohort. Teams that defer it to next quarter end up making pricing decisions from vibes for eighteen months and discovering, after the fact, that a single customer segment was responsible for half the inference bill at negative margins.

The most dangerous failure in agent systems is not the loud one. It is the agent that confidently declares "Task complete" and returns a polished summary of work it never did. The file was never written. The webhook never fired. The database row is still the way it was an hour ago. But the trace is green, the completion counter ticks up, and the dashboard tells leadership the new feature is working.

This is the hallucinated success problem, and it is the single hardest bug class to catch in production because it evades every cheap signal you have. The agent did not crash. It did not time out. It did not return an error. It narrated a plausible, coherent, and completely fabricated account of a successful execution. Your observability stack was built to catch noisy failures. Silent success looks identical to real success until a user notices the output is wrong.

Teams add human approval to an AI workflow the same way they add if (isDangerous) requireHumanApproval() to a codebase: as a binary switch, checked once at design time, then forgotten. The metric on the architecture diagram is a green checkmark next to "human oversight." The metric that actually matters — how long the human took, whether they read anything, whether the item was still relevant by the time they clicked approve — rarely has a dashboard.

Treat the human approver as a binary switch and you have built a queue without knowing it. And queues have dynamics: backlog that grows faster than you staff, staleness that makes yesterday's decision meaningless, fatigue that turns review into rubber-stamping, and priority inversion that parks the one decision that mattered behind three hundred that didn't. None of this is visible in the architecture diagram. All of it shows up in the incident retro.

Your dashboards are clean. The deployment from yesterday rolled back cleanly. The model version is pinned. The prompt didn't change. But your TTFT p99 just doubled, your customer success channel is on fire, and the only honest answer you can give is "it's the provider." That answer feels small — like a shrug — and it usually leads to a follow-up question that nobody on your team can answer: prove it.

This is the part of hosted LLM inference that the marketing pages do not discuss. When you call a frontier model API, you are sharing a GPU, a PCIe fabric, a continuous batch, and a KV-cache budget with workloads you cannot see. Your p99 is a function of their bursts. The economics of large-scale inference depend on multiplexing tenants tightly enough that hardware utilization stays north of 60-70%, which means your tail latency is structurally coupled to the largest, jankiest, lumpiest tenant on the same shard. You are not buying capacity; you are buying a slice of a queue that someone else is also standing in.

Open any 2024-era AI feature's trace and the model call is the whale. Eight hundred milliseconds of generation surrounded by a thin crust of retrieval, auth, and a database lookup rounding to nothing. Every architecture decision that year — the caching, the prefetching, the streaming UX — was designed around hiding that whale.

Now pull the same trace for the same feature running on a 2026 inference stack. The whale is a dolphin. A cached prefill returns the first token in 180ms. Decode streams at 120 tokens per second. The model is no longer the slow node. Your own infrastructure is, and most of it hasn't noticed.

This reordering is the most important performance shift of the year, and it's the one teams keep under-reacting to. The p99 floor on an AI request is now set by the feature store call, the auth middleware, and the Postgres lookup that was always that slow — nobody just cared when the model was taking nine-tenths of the budget.

A regression suite that flips green-to-red without a single prompt change is usually one of three things: a broken test harness, a flaky retrieval store, or a judge that learned new taste over the weekend. The third one is the most common and the least debugged, because no commit in your repo caused it. The scoring model got a silent quality refresh, and every score you compare against last month's dashboard is now denominated in a different currency.

This is the uncomfortable part of LLM-as-judge: you have two moving models, not one. The candidate model is the thing you ship; the judge model is the thing that tells you how the candidate is doing. When both evolve independently, score deltas stop meaning what they used to, and the dashboard that your PM refreshes every morning quietly lies.