89% of Agent Teams Can Watch. Only 52% Can Test.
89% of teams with AI agents in production can observe them. Only 52% can evaluate them. The 37-point gap is where agents go to die.
The 37-point gap between agent observability and agent evals.
That's the number buried in LangChain's State of Agent Engineering 2026 report. 1,300 teams surveyed. 57% have agents in production. 89% of those teams have observability for their agents — logs, traces, dashboards. Only 52% have evals. Offline eval coverage sits at 52%. Online monitoring sits at 37%.
The infrastructure pyramid looks like this:
- Bottom layer: observability. Logs, traces, token counts, latency. Almost everyone has it.
- Middle layer: offline evals. Golden datasets, regression thresholds, CI gates. Roughly half.
- Top layer: online monitoring + evaluation. Live shadow traffic, score regression detection, automatic rollback. About a third.
Most teams stopped at the bottom and shipped anyway. The bugs that kill production agents don't show up in traces. They show up in subtle behavior drift across hundreds of runs.
Andrej Karpathy published a 12-page PDF on August 3 that names the missing piece: graph engineering.
His argument is straightforward. An agent loop runs 700 experiments per session, makes 700 separate decisions, and remembers none of them. A graph remembers everything. State, context, prior tool calls, accumulated errors — all persistent, all queryable, all replayable. That's not a database. It's a new primitive for agents that need to think across more than one turn.
Three shifts to make this quarter:
-
Stop treating observability as the finish line. If your agent logs work but your eval suite doesn't run on every PR, you have a logged experiment, not a production agent.
-
Move from prompt eval to graph eval. Test the state transitions, not just the final outputs. Did the agent remember what it learned in turn 3 when it reached turn 7?
-
Treat the graph as a first-class artifact. Version it. Replay it. Branch it. The teams shipping agents that improve over time are the ones comparing graph snapshots across releases.
The frontier in agent engineering isn't a better model. It's the graph beneath the loop.
Key Takeaways
- LangChain's 2026 State of Agent Engineering report (1,300 teams surveyed) found 89% of teams with production agents have observability but only 52% have evals — a 37-point gap that defines the production-agent failure mode
- The infrastructure pyramid is observability → offline evals → online monitoring. Most teams ship with only the bottom layer and call it done
- Andrej Karpathy's 12-page 'graph engineering' PDF (Aug 3) names the missing primitive: an agent loop runs ~700 experiments per session and forgets them all, while a graph retains state, context, prior tool calls, and accumulated errors
- The bugs that kill production agents don't show up in traces — they show up as subtle behavior drift across hundreds of runs. Logs are necessary, not sufficient
- The frontier in agent engineering isn't a better model. It's the graph beneath the loop
Sources
Get the next briefing
Signal-first AI briefings, weekday mornings.
One concise briefing with three signals, why they matter, and one action to take.
Free. No spam. Unsubscribe anytime. · Weekday mornings.
Share this article