Where did all your tokens go?
Annie Freeman
Developer Advocate, Coralogix
"This feels faster." "That seems better."
"This feels faster." "That seems better."
Vibes
Measure coding agent impact using OpenTelemetry to improve efficiency
Hello, I’m Annie
Developer Advocate at Coralogix
Platform Engineer at Xero
Trained as a lawyer in New Zealand
1
2
3
4
Coding agent usage today
Coding Agent Instrumentation with OTel
My experiment
Takeaways
03 / 27
04 / 27
04 / 27
Every token is energy…
https://simonpcouch.com/blog/2026-01-20-cc-impact/
Every token is energy…
… and the creation of software now contributes to that energy usage
Coding agent instrumentation with OpenTelemetry
The data your agent is already emitting
You instrument nothing.
Easy set up
THE EXPERIMENT
Default open stance
Default open stance
TDD
Skip permissions
THE RESULTS
Where did this money go?
Caching in code agents
But why?
It’s not just me
We then analysed across
21,329
sessions.
AND ALMOST NO ONE IS LOOKING
52%
name missing metrics as their single biggest challenge with AI.
Only 29% measure impact at all, and mostly on sentiment.
LeadDev · State of AI-Driven Software Releases 2026
coralogix
08 / 27
ONE MODEL FOR EVERYTHING AFTER THIS
cost = context × turns × model rate
Three multiplicative levers. And it's the energy model too, because those same three things drive the compute.
coralogix
11 / 27
Four levers to be more token efficient
1
2
3
4
Code creation contributes to energy usage, so we also need to think about efficiency here
It’s super easy to set up telemetry collection!
The cache hit rate is a key lever that impacts token usage
You can connect this data to DORA metrics to decide what approaches work best for you
KEY TAKEAWAYS
04 / 27
Let’s connect
52
52
Annie Freeman