● SAP TechEd Berlin 2026 · Community talk CT1106

Zero token architecture for Joule Agents

Tue 27 Oct 2026 14:00–14:20 CET Berlin + Add to calendar

edgar-martinez.dev/teched — agent session

❯ session start --talk CT1106 ✓ loaded 3 types of run · 4 steps for Monday · 1 repo ℹ this page is a Type 1 run: pre-written answers, zero tokens, no model at runtime. ✓ ready. Scroll to continue, or type a question below.

you

Quick recap: what was the idea again?

edgar · agent

readct1106.md20 minutes · 3 types of run

render<RunLadder />

An agent solves a new kind of case once, with a model, and the solution becomes code or configuration that gets validated. The next time that case shows up, it runs with no model at runtime and no tokens spent.

<RunLadder types={3} />rendering

TYPE 3
The agent explores the whole case and decides everything. Every new kind of case starts here.
10k–50k tokens
TYPE 2
A playbook runs the fixed steps. The model is called only to interpret what it sees.
1k–5k tokens
TYPE 1
Rules and code run the whole case. No model at runtime. This is the ghost.
0 tokens

Every run of every agent lands on one of them. The mix is the metric: the point is not zero everywhere, it is knowing which runs still need a model.

follow up

you

How would I try this with our own agents?

edgar · agent

readstart-monday.md4 steps · no new platform

render<Plan steps={4} />

Start Monday. Four steps, no new platform: you need logs and one afternoon.

<Plan when="Monday" />rendering

01Record every run.Input, steps taken, tokens used, outcome. If it is not logged, it did not happen.
02Measure your Type 1:2:3 baseline.Count how many runs fall into each type today.
03Shortlist repeated paths.Rank them by repeats × tokens. The top of the list is where the money is.
04Freeze one.Turn it into code or configuration, keep a fallback to the agent, and get a human signature before it ships.

Steps 2 and 3 are what the repo's calculator does with your own log. Run the baseline and want a second pair of eyes? Email me the numbers.

follow upEmail me the numbers

you

Is there code I can try?

edgar · agent

readgithub.com/eamarce/zero-token-joule-agentsv0.2.0 · SAP CAP + LangGraph.js · MIT

render<RepoCard />

Yes. I wrote an open-source agent inspired by the talk, and it runs on your laptop with no API key and no SAP account.

<RepoCard license="MIT" />rendering

zero-token-joule-agents

SAP CAP + LangGraph.jsNode 22+MIT

An agent that stops calling the model for the cases it has already learned, and measures how much of its traffic still needs one.

The cycle the talk describes, on 40 sample supplier emails: explore, propose code, a person signs, then 0 tokens until a re-check demotes it.
A calculator for your own agent's log (run-log JSONL or OpenTelemetry traces): your Type 1:2:3 ratio and the paths worth signing.
Callable from SAP Joule over A2A.

$ git clone https://github.com/eamarce/zero-token-joule-agents
$ npm install && npm start   # no API key, no SAP account

Open the repo →

follow up

you

Could we talk about our agents?

edgar · agent

readspeaker.mdCogent IBS · 17+ years

render<Speaker />

Gladly. I work on SAP BTP and AI, taking agents from demo toward production, so bring one you already run.

<Speaker />rendering

Edgar Martinez

Cogent IBS · SAP BTP architecture and delivery

17+ years in enterprise software, working remote from Mexico with teams in the US and Europe. ASUG speaker.

Book a call

25 minutes, one agent you already run.

We look at where its runs land today and what you could freeze first.

Book 25 minutes

More of my work

The rest lives on edgar-martinez.dev.

Six client problems, as they were told to me
15+ systems in production since 2008
How to work together

Visit edgar-martinez.dev ↗

follow upConnect on LinkedInVisit edgar-martinez.dev