Overview
Knowledge hierarchies escalate exceptions upward so scarce expertise handles hard cases (Garicano, 2000). Generative AI changes that calculus: automation versus augmentation shifts who faces routine work and who absorbs exceptions (recent organisation-theory models of GenAI in knowledge economies). In engineering, multi-agent systems add relays that act as information bottlenecks — helpful when they remove noise, harmful when they drop task-critical context — while governance layers such as the Organizational Control Layer (OCL) separate proposal generation from environment-facing execution. What is missing is a topographic experiment: encode escalation as a conductance constraint, measure whether the judgement junction is actually exercised, and quantify agentic debt when institutional escalation paths exist on paper but are bypassed in the realised communication graph (parallel channels, aggressive summarisation, prompt injection).
Prerequisite
L172 Information, Energy and Intelligence (IEI), or equivalent preparation in information theory, maximum entropy, and information geometry.
FAQs
- What are the prerequisites?
L172 Information, Energy and Intelligence (IEI), or equivalent preparation in information theory, maximum entropy, and information geometry.
- What will I learn in this Project?
You will learn knowledge-hierarchy models of organisations (Garicano), information-bottleneck accounts of multi-agent relay compression, and execution-boundary governance (external control layers that approve, revise, block, or escalate). You will implement adversarial evaluations of escalation bypass and define measurable agentic debt.
- What is the objective of the project?
(1) Build a knowledge-hierarchy task with routine versus exception difficulty, junior agents, and an expert / human-simulating escalation target. (2) Encode escalation thresholds and an external control layer (approve / revise / block / escalate) separate from the LLM proposer. (3) Adversarially induce bypass (prompt injection, side channels, summaries that drop exceptions) and measure (i) task performance, (ii) fraction of exceptions that hit the judgement layer, (iii) debt = institutional escalation path − realised topography. (4) Compare to token-economy pruning (e.g. AgentPrune-style sparsity): when does saving tokens destroy escalation? Deliverables: experimental platform, adversarial suite, debt metrics, thesis chapter.
- How does this fit into the bigger picture?
Agentic debt and the judgement layer are central ideas in the organisational study of agentic systems. This project is an interventional approach: the organisational-design vocabulary (escalation thresholds, requisite-variety-matched channels) would be implemented as conductance structure, with failure modes that WP3 can test in real institutions. Co-supervised with the Interfaces / DOAgent line of work.
