BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.adass.org//adass2026//talk//FBQB9X
BEGIN:VTIMEZONE
TZID:AWST
BEGIN:STANDARD
DTSTART:20000101T000000
RRULE:FREQ=YEARLY;BYMONTH=1;UNTIL=20051231T160000Z
TZNAME:AWST
TZOFFSETFROM:+0800
TZOFFSETTO:+0800
END:STANDARD
BEGIN:STANDARD
DTSTART:20070325T040000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:AWST
TZOFFSETFROM:+0900
TZOFFSETTO:+0800
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20071028T030000
RRULE:FREQ=YEARLY;BYDAY=4SU;BYMONTH=10;UNTIL=20081025T190000Z
TZNAME:AWDT
TZOFFSETFROM:+0800
TZOFFSETTO:+0900
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-adass2026-FBQB9X@pretalx.adass.org
DTSTART;TZID=AWST:20261103T090000
DTEND;TZID=AWST:20261103T093000
DESCRIPTION:Modern AI tools have made code cheap to produce. Trust in the r
 esults \nhas not become cheap\, and in science and critical systems it is 
 essential. Without \ntrust\, the speed that AI offers is not an advantage.
 \n\nI lead a small team building the Data Center for the CCAT Observatory\
 , \nincluding the complete data life cycle after observations\, from the \
 noperations database and archive through the analysis workflows\,\nHPC inf
 rastructure and data access frontend. We were early adopters of \nAI tools
  and moved from consulting chat systems about implementation \ndetails\, t
 o discussing architecture\, to handing coding work completely \nover to ag
 entic AI. With a team size of two FTEs\, this let us build more \nof what 
 I had envisioned for the data center\, earlier and to a higher \nstandard.
  One example was a rewrite of our calibration software into a \nworking\, 
 parallelized Rust implementation in one week\, with parity \nagainst its p
 redecessor and a substantial speedup. But our team was \nskeptical this ne
 w code really calibrates the data reliably. Before \nagentic AI tools the 
 slow development and frequent discussions built \ntrust in the code\, for 
 both the developer and the team. Now\, a fast \nrewrite happens on one per
 son's computer\, and without precautions leaves \ncolleagues behind\, and 
 oneself with a tool of uncertain behavior on an \nunfamiliar code base.\n\
 nI will present the workflow I built for us to reach that trust without \n
 giving up development speed. Its purpose is to keep humans and agents \nwo
 rking from the same understanding. We let the AI agents use \nprogramming 
 languages and frameworks that we know well. Where exact \nresults matter w
 e create deterministic testing harnesses that we can \naudit. Because AI s
 essions are ephemeral\, I designed the workflow so \nthat machine-readable
  definitions force each session to leave clear \nartifacts in well defined
  places: implementation plans\, decision \nrecords\, and documentation. Th
 ese artifacts inform and onboard \ncolleagues\, give the next agent its co
 ntext\, and let us later understand \nwhat we built and why. However\, we 
 noted that artifacts that satisfy the \nagent are often unreadable for hum
 ans. We therefore invest in shaping \ntheir language and aligning on a sha
 red\, defined vocabulary\, so that the \ndocuments can be understood witho
 ut the context of the sessions or \nnative-level English. I will show what
  has worked for us\, where it has \nnot\, and what we are still doing abou
 t it. \n\nLooking ahead\, I am unsure where agentic AI will leave us as de
 velopers \nand scientists. The calibration rewrite was fast\, but only bec
 ause I understood the old \nsoftware well after working with it for 10 yea
 rs and we had a test suite. Work has been \noffloaded before\, for example
  we trusted libraries to do complex jobs\, \nhowever this is the first tim
 e that judgement can seemingly be offloaded \nto AI. Before\, this judgeme
 nt could only be built by study\, failure and \ndiscussion. This leads to 
 one central question I have: how does the next \ngeneration acquire that j
 udgement and experience if they never have to \nearn it first?  Or do we 
 hand judgement over altogether\, willingly or not?\n\nI am convinced that 
 using AI agents for software engineering can improve \nthe science\, and t
 he support of it\, tremendously\, but only if we keep \npeople and our own
  judgement in the loop.
DTSTAMP:20261015T133415Z
LOCATION:Banquet Hall
SUMMARY:Code Is Cheap\, Judgement Is Not: Building Software and Systems We 
 Can Trust with Agentic AI - Christof Buchbender
URL:https://pretalx.adass.org/adass2026/talk/FBQB9X/
END:VEVENT
END:VCALENDAR