The leader track 🎯 · for the people who own the on-call rota · 6 x 45 min · no tools required
Judgement mode: why a green dashboard held through a real incident, what to instrument before you know what you are looking for, which dashboards answer questions and which merely glow, what AI actually costs you once you count failed attempts, why static alert thresholds produce the same fatigue that made alarm hazards the top health-technology hazard two years running, and whether to buy, build or stay OTel-native.
The trace is the receipt
A green dashboard held through a real incident while executives read stale numbers. Why "it works in evals" and "it worked at 09:14" are different claims, and what the second one costs.
🎛️ Leader 2 · easyWhat to instrument before you need it
Coverage is a decision made before you know what you are looking for. You get the model calls free, and the model calls are rarely the bug.
📈 Leader 3 · mediumDashboards that answer questions
Three questions a trace must answer, why an aggregate p95 hid an intermittent bug, and why cardinality forces metrics and traces to do different jobs.
💸 Leader 4 · mediumWhat AI actually costs you
Cost per call is the wrong denominator and it moves the wrong way. Caching economics, the tokenizer that added 30% for free, and cost per successful task.
🚨 Leader 5 · decidingOn-call for AI
Alarm hazards topped the health-technology hazard list two years running, and 88.8% of ICU arrhythmia alarms were false. Your Slack channel is next.
🧭 Leader 6 · decidingBuy, build, or stay OTel-native
Lock-in is an attribute-dialect problem. The market splits three ways, the licences have teeth, and one tool captures your users prompts by default.
The builder track 🛠️ · for AI engineers, platform engineers & technical PMs · 10 x 45 min · hands-on
Craft mode: spans and traces from first principles, the current GenAI conventions including the names that changed under you, instrumenting a real agent loop, cost and latency attribution that survives it, a live root-cause hunt through a broken trace, the empirical failure taxonomy, sampling and what observability costs, burn-rate alerting, drift you did not deploy, and the privacy split that keeps you debuggable and compliant at once.
Spans and traces
Six words the course runs on, why a log line stops working the moment there is a loop, and the most useful number you can compute today: unaccounted time.
📗 Builder 2 · easyOTel GenAI conventions
The exact current attribute names with their requirement levels - including the one that no longer exists, and the five events that collapsed into one.
🕸️ Builder 3 · mediumInstrumenting an agent
The loop is width, not depth. Why plan owns its model call while tools are siblings, and why upstream ships no instrumentation for the framework you use.
⏱️ Builder 4 · mediumCost and latency per span
Attribution that survives an agent loop, the double-counting trap, and the one ratio that tells you a faster model will not help.
🔍 Builder 5 · hands-onReading a broken trace
The live hunt. Eight coverage rungs, 32.4% of the wall clock to 100%, zero of five questions to five - and one lever that changes nothing but the bill.
🗂️ Builder 6 · hands-onThe failure taxonomy
Fourteen measured failure modes from 150 annotated traces. The biggest bucket is not "the model was dumb" - it is specification and design, at 44.4%.
🎚️ Builder 7 · hands-onSampling and its own cost
You want the slow traces, and head sampling cannot give them to you. Tail sampling, the 30-second default that breaks on a 20-minute agent, and cardinality.
🔔 Builder 8 · hands-onAlerting and SLOs
Burn rate instead of thresholds. Five reasons static limits misfire on LLM systems, and the alerts that fire at HTTP 200.
🌊 Builder 9 · hardestDrift and silent changes
Four kinds of drift, three ways a provider changes the model under you, and the embedding mismatch that degrades retrieval to noise while every span stays green.
🔐 Builder 10 · hardestPrivacy and retention
Metadata at one hundred percent, content on a sample. The split that keeps you debuggable and compliant, and the model choice that overrides your policy.
Choose your path 🗺️
Four doors into the same conviction: an un-instrumented step cannot fail, as far as your trace is concerned.
The knowledge map 🧠
Both tracks at a glance - hover a session to spotlight its concepts, click any node to jump in.