BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.com//pydata-amsterdam2026//speaker//MK9CY8
BEGIN:VTIMEZONE
TZID:CET
BEGIN:STANDARD
DTSTART:20001029T040000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20000326T030000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-pydata-amsterdam2026-KRSRVM@pretalx.com
DTSTART;TZID=CET:20260911T141000
DTEND;TZID=CET:20260911T144000
DESCRIPTION:LLM-based coding agents can navigate codebases\, edit files\, r
 un tests\, and iterate with little supervision. Teams have started applyin
 g them to ML projects\, where the volume of repetitive experimental work m
 akes automation appealing.\n\nHowever\, coding execution and ML optimizati
 on are not the same problem. In software engineering\, success is usually 
 local and binary. In ML\, code correctness is only a prerequisite: progres
 s is measured across repeated experiments\, and getting there requires exp
 loration over many trajectories rather than a single one. Coding agents\, 
 by contrast\, are built to push one trajectory to completion. Over long ho
 rizons they drift. Changes get bundled into one step\, so results cannot b
 e attributed. Context is lost between sessions\, and hypotheses that were 
 already tested come back in slightly different wording.\n\nAdding a higher
 -level LLM agent on top to orchestrate the others does not fix the problem
 . The orchestrating agent drifts too. Karpathy's [`autoresearch`](https://
 github.com/karpathy/autoresearch) is a recent illustration: a coding agent
  given an `.md` prompt spins up its own worktrees and logs experiments to 
 a CSV\, but the trajectory it explores stays single and shallow. Stable be
 havior across hundreds of experiments requires deterministic code\, not be
 tter prompts.\n\nThis talk shows how to build such a system. A non-LLM orc
 hestrator runs experimentation as a tree search over hypothesis-constraine
 d modifications\, with each branch living in its own git worktree. Differe
 nt node types separate proposing a change from implementing it and from de
 bugging it. Results flow into a real experiment tracking system through a 
 token-efficient query interface\, so an agent's context can be reset betwe
 en calls without losing history. The orchestrator dispatches to existing c
 oding CLIs (e.g. `Claude Code`\, `Codex`) without abstracting them away.
DTSTAMP:20260710T150511Z
LOCATION:Room 2 (350)
SUMMARY:Deterministic Orchestration for ML Experiments with Coding Agents -
  Oleh Kostromin\, Iryna Kondrashchenko
URL:https://pretalx.com/pydata-amsterdam2026/talk/KRSRVM/
END:VEVENT
END:VCALENDAR