Responsibility
Who can trigger an action? With which resources, limits and human approval?
Research
Understanding autonomy, programming responsibility
Agents can choose actions and interact with real systems. Our research studies how to give them explicit, verifiable boundaries that developers and users can understand.
The questions we study
Composition
How can agents delegate while keeping responsibilities and permissions clear?
Evidence
How can an action be linked to its authorization and traced back through the program?
From principles to mechanisms
Probabilistic Agentic Programming (PAP)
PAP puts explicit responsibility at the center of agentic programming. Its architectural principles are independent of a particular language’s syntax.
Formalization
A formal model describes authorization and mediation conditions. Its proven properties establish what these mechanisms guarantee within the studied fragment.
Touba
The language and runtime make these principles executable through compiled contracts, checks before effects and traces linked to source code.
What the results tell us
Answer quality and authority are separate questions
A model may propose an incorrect action. Effective governance keeps execution within the program’s limits and provides an inspectable trace of its decision.
The interface matters
How capabilities and history are presented influences model outcomes. We evaluate this alongside execution mechanisms.
Contracts should remain clear across layers
Connecting source, compiled representation and execution supports the study of responsibility across the whole chain, including delegation.
Papers
- arXiv preprint
Responsibility Boundaries for Agentic Systems: A Comparative Evaluation of Probabilistic Agentic Programming
Responsibility boundaries: formalization, abstractions and comparative evaluation.
Read on arXiv
Challenges ahead
We study how to extend formal properties to concurrency and interruptions, govern remote effects and help developers understand and audit responsibilities.
Timeline
- Fall 2025
The project begins with explicit responsibility and bounded autonomy.
- Winter 2026
Review of prior research and formalization of PAP foundations.
- Summer 2026
Work on responsibility boundaries and comparative evaluation.
- Fall 2026
Further formalization and development of Touba’s executable mechanisms.