Agentic Coding Summit #2: Where Did All My Tokens Go?

Welcome & Introduction

Opening remarks for the second Agentic Coding Summit.

Where Did All My Credits Go? Understanding and Reducing Token Usage in GitHub Copilot

Since June 1, every Copilot interaction is billed by tokens, and the internet is full of tricks that promise 60, 70, even 95 percent savings. But which of them hold up? In this talk, I explain how tokens and AI credits actually work, show how to measure your own usage with tools you already have, and put several popular reduction techniques to the test on real tasks from my own repositories — with honest numbers instead of headline percentages. You'll leave knowing where your credits go and how to check any savings claim yourself.

Dr. Matthias Liebeck · Watch recording

Copilot, Traced. Your AI Pair Programmer Is a Distributed System. Instrument It Like One.

Abstract: You have instrumented every service you own — traces, dashboards, alerts, an on-call rota — and then you let the tool that writes your code run as a black box. How many tokens did your last Copilot turn send? What did it cost? Was any of it cached, or did you pay full price to re-send a prompt the model had already seen thirty seconds earlier? Almost nobody can answer, and almost nobody realises that Copilot has been emitting OpenTelemetry this whole time — real OTLP, real GenAI semantic conventions, four settings away from landing in the Grafana or Application Insights you already run. This talk turns Copilot into a service you can actually observe: we read one real turn as a span tree, learn why a single moved byte costs twenty-one thousand tokens, build a telemetry pipeline live and take the generated config with us, and put a number on what a bad prompt prefix costs a forty-person team per month. We also cover the parts the docs get wrong — the cost attribute that is documented under one name and shipped under another, and why "content capture off" is not the same as anonymous. You'll leave able to see your own prompt-cache hit rate in about twenty minutes, on your laptop, without asking anyone's permission.

Maxim Salnikov · Watch recording

Break

Copilot Already Optimizes Automatically. What Is Left for You?

Abstract: Since June 1, Copilot usage has been billed by token consumption in AI Credits. At the same time, the harness already optimizes large parts of the context. This talk shows what happens automatically — and which effective levers remain in your hands.

Oliver Dzaeck · Watch recording

Less Token Waste, Same Delivery: A Practical Copilot Workflow Experiment

Abstract: This talk presents a practical approach to reducing GitHub Copilot credit usage in everyday engineering work.

I walk through one task completed in multiple runs with identical model settings and show which workflow changes made the difference. The implementation goal stays the same, while the overall workflow becomes leaner and more cost-aware.

You will leave with a compact playbook you can apply immediately. It covers where token waste typically appears during coding and which lightweight habits reduce it.

Felix Burkhard · Watch recording

Roundtable: Token-Based Billing: What Actually Changed?

All speakers discuss what changed in daily work, which savings techniques survived contact with reality, and what they want GitHub to fix.

Wrap-up

End