HACKATHONEfficient Local Agents · IBM Vienna
Small model.
Smart harness.
Same models, same budget for everyone: a fixed small language model plus Jev by TypeSafe for instant structured decisions. Build the most reliable and cost-efficient autonomous agent: your engineering decides who wins. The scenario is revealed at the opening.
NOV
14
2026
- When
- Sat, 14 Nov
- 09:00 – 19:30+
- Where
- IBM Vienna
- Obere Donaustraße 95
- Teams
- Up to 4
- or join solo
- To win
- IBM credits & goodies
- Certificates for winners and all participants
The setup
- A fixed small language model as the main model, the same for every team (announced at the opening)
- Jev 1.14 (
typesafe/jev-1.14) for cheap typed decisions - Both only via OpenRouter, with your team's keys
- Your harness runs on your laptop and talks to our Task API
Scoring
50% accuracy, 50% cost, coupled: final = A · (0.5 + 0.5 · C). Accuracy counts correct answers, correct citations and safe actions. Cost is your OpenRouter spend on a log scale between fixed bounds.
The day
- 09:15 challenge reveal, dev scenario opens
- →17:00 unlimited dev runs with instant feedback
- 17:00 secret final scenario: one unattended 30-min run
- 18:00 leaderboard reveal & awards
Task families
Information synthesis
Which sites currently miss a required standard?
Conflict resolution
Two reports disagree about a key fact. What can safely be concluded?
Resource allocation
Allocate a limited stock across sites under given constraints.
Tool use
Find feasible options with the tools after something changed.
Dynamic replanning
A site becomes unavailable; update the plan.
Verification & security
Spot unsupported claims, refuse injected instructions, don't over-refuse.
In collaboration with
