Agents that are boring in production.
I build AI agents for logistics and customer support teams in Berlin and Hamburg. Every claim on this page sits level with its evidence.

E0.1 Status16 Nov 2026
3 agentsin production, all with a person who can stop them
E0.2 Last incident14 Mar 2026
Carrier API timeout. The dispatch agent paused itself and paged a dispatcher in 40 s. No message went out wrong.
E0.3 AvailabilityCET
Open to contract work from February 2027, three days a week.
A dispatch agent that asks before it promises
Kranich Freight, Hamburg. Logistics, 2024 to now.
It reads tracking, drafts the delay message to the customer, and waits for a dispatcher on anything worth more than €2,000.
41% of delay emails now go out without edits. Calls about late freight fell by 28%.

E1.1 Trace, run 7f3a2 Oct 2026
06:12:04 read tracking KF-20418 84 ms06:12:04 check delay over 4 h 2 ms06:12:05 draft customer message 690 ms06:12:05 hold value €3,180 waiting06:14:37 ok M. Sommer, dispatch06:14:37 send 2 recipients 112 ms
E1.2 Evals, v31
412 / 430cases pass before release; any failure blocks it
E1.3 Holds
0messages over €2,000 sent without a person in 90 days
Support answers that show their source
Nordlicht Energie, Berlin. Customer support, 2025 to now.
Billing answers arrive with the policy paragraph attached. Anything about disconnection is handed to a person, every time.
First replies take 38 seconds, and escalations dropped by 23% in the first quarter.

E2.1 Evals, billing v14600 tickets
answer cites a policy paragraph 97.2%disconnection handed to a person 48 / 48refund promised without approval 0median first reply 38 s
E2.2 Incident note22 Jun 2026
A tariff changed overnight. Until the policy index rebuilt at 07:10, the agent said it would check with a colleague. Nobody was quoted the old price.
Evals you can run before every deploy
Probelauf, open source under the MIT licence.
Probelauf replays real agent traces as tests. It runs in CI and fails the build when a tool call or a promise changes.
2,100 stars, 46 contributors, and a talk at the Spree ML meetup in September.

E3.1 CI run 1184main
$ probelauf run traces/ --against mainreplaying 214 tracespass 209 same tool calls, same promisesfail 3 refund promised, no approvalfail 2 wrong carrier endpointbuild failed: 5 regressions
E3.2 Why it exists2024
A prompt change once made a support agent promise refunds. Unit tests passed. Replaying last week's traces would have caught it.
Nine years from search ranking to agents
I started on ranking models, moved to retrieval, and have built agents since 2023. I write about traces and talk about boring systems.

E4.1 Roles2017 to now
2023 to now independent, agents2020 to 2023 staff ML, Wegweiser search2017 to 2020 ML engineer, Fundus ranking2017 MSc Computer Science, TU Berlin
E4.2 Writing and talks3 recent
Agents that are boring in production, Spree ML 2026. Read the trace first, essay 2025. Approval is a feature, workshop 2024.
Ask about the work, see the trace
Each answer shows what it looked at, what it matched and where the evidence sits on this page.
Enter asks. Answers come from Lena's notes on this page.
Answers from Lena's notes. Scripted, not Lena.
The trace for your question appears here: what it looked at, what it matched, and the answer with its evidence row.