BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:AI Center Seminar - AI Fundamentals series - Dr. Laurent Bindschae
 dler (Max Planck Institute for Software Systems)
DTSTART:20261124T143000
DTEND:20261124T153000
DTSTAMP:20261109T192700Z
UID:b622522fdd046cb36c1cd8f787ebd9f617b8fdc9e1111858f8eecbbb
CATEGORIES:Conferences - Seminars
DESCRIPTION:Dr. Laurent Bindschaedler (Max Planck Institute for Software S
 ystems)\nThe talk is organized by the EPFL AI Center and the Laboratory
  for Intelligent Global Health and Humanitarian Response Technologies (LIG
 HT) as part of the AI Fundamentals seminar series.\n\nHosting Professor: P
 rof. Mary-Anne Hartley\n\nTitle\nA Systems Approach to Safe and Reliable g
 ents: Four Contracts for Agent Runtimes\n\nAbstract\nA computer use agent 
 is asked to place an order. The request times out after the supplier has a
 lready accepted it. The agent retries\, two orders ship\, and most benchma
 rks today score this a success. Every step was reasonable\, so it is genui
 nely hard to argue that the model made a mistake. And yet the outcome is c
 learly wrong.\n\nThe usual answer to failures like these is training a bet
 ter model to lower the failure rate. However\, lowering a rate and making 
 a guarantee are not the same thing. A number of applications cannot be bui
 lt just on the failure rate. This talk starts from a different question: w
 hat should the system around the model be required to guarantee instead?\n
 \nOver the past two years my group has been working on four contracts\, ea
 ch stating guarantees the system must keep. First\, effect\, which says wh
 en an action becomes real and whether it can happen twice. Second\, state\
 , which says what an agent must not forget and what must happen when it do
 es. Third\, composition\, which asks whether guarantees that hold for one 
 agent still hold once several agents and many steps are involved. Finally\
 , intent\, which asks what plays the role of source code when agents gener
 ate and execute it on the fly. The four sit at different stages of maturit
 y.\n\nI will present them alongside the mechanisms we have been developing
 : typed agent memory that raises a fault instead of forgetting silently\, 
 a mediator that admits an effect only when the specification allows it\, a
 nd a tamper-evident record of what an agent actually did. Writing the cont
 racts down is a first step toward a shared standard\, a kind of POSIX for 
 agents\, though it is not the last. Systems in use today almost certainly 
 violate these properties: retries that the supplier sees as new orders\, r
 ules dropped at compaction with no fault raised\, and approvals reused aft
 er the reviewed code has changed. The next steps will be working out what 
 to enforce and refining the properties as we build more systems.\n\nMy goa
 l for this talk is to convince you that training agents to be safe and rel
 iable will need a complementary systems approach.\n\nBio\nLaurent Bindscha
 edler leads the Data Systems Group at the Max Planck Institute for Softwar
 e Systems and is an Associate Fellow at Saarland University. He builds sys
 tems for AI agents\, with current work on agent runtimes\, trustworthy cod
 e generation\, efficient LLM serving\, and data systems for AI. He holds a
  PhD from EPFL and was a postdoctoral fellow at MIT CSAIL\, with earlier w
 ork on operating systems\, databases\, and distributed systems. His work h
 as appeared at SOSP\, ASPLOS\, EuroSys\, ICML\, EMNLP\, and NDSS.
LOCATION:ELE 117 https://plan.epfl.ch/?room==ELE%20117 https://epfl.zoom.u
 s/j/65942397298?pwd=mH0XCHmFp3EAGsGQZ3LZpZlQFpA9GD.1
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR