A research notebook · revised as the questions evolve
Research.
A capable agent is one thing. Understanding what happens between agents—and across time—is another.
I study the mechanisms behind system-level behavior in language-model agents, drawing on interaction traces, network science, security analysis, and simulation.
From a question to a mechanism.
NOTE 01 / INTERDEPENDENCE & SECURITY
How does a local failure become a system-level failure?
Information crossing from one agent to another can change its role: evidence may become instruction, and a local error may be relayed or amplified downstream. Shared state and external tools create further routes across trust boundaries. I investigate how such dependencies affect system-level security and reliable coordination.
I examine temporally ordered interaction traces, reconstruct cross-agent dependencies, and distinguish communication-, memory-, and tool-mediated influence. The goal is to identify origins and propagation pathways and study interventions that preserve useful collaboration.
OPEN QUESTION Which local signals anticipate system-level compromise, and which dependencies matter for attribution and containment?
NOTE 02 / CHANGE & OVERSIGHT
If an agent evolves, what else must evolve with it?
An agent can change substantially without changing its underlying model. New memories, tools, skills, or workflows alter the actions it can take and the assumptions under which it was previously evaluated. I study these updates as changes in the agent's operational state, not only as changes in task performance.
I compare agent states before and after an update, characterize changes in available actions and dependencies, and investigate how oversight can respond when new capabilities introduce new risk.
OPEN QUESTION What can the changes between successive agent states tell us about risks that have not yet appeared in observed behavior?
NOTE 03 / SCALE & COLLECTIVE DYNAMICS
When do local interactions become collective behavior?
Collective behavior depends on who interacts with whom, how often, and under what constraints. Even when individual agents follow similar instructions, their interaction networks can develop uneven influence, specialization, bottlenecks, and extreme coordination events.
I measure coordination events across system sizes and network topologies, study the distribution of influence and rare interactions, and use generative agent-based models to connect local decisions with population-level outcomes.
OPEN QUESTION Which collective patterns persist across tasks and network structures, and when does increasing the number of agents worsen coordination?
NOTE 04 / OBSERVABILITY & EVALUATION
What should we measure when the final answer looks right?
A correct final answer does not establish that the decisions leading to it were reliable. I study what an agent observed, which evidence it used, how its commitments changed, and what happened inside the tools, memory, and orchestration layers surrounding it.
I use trace-level monitoring, attribution, and intervention alongside decision-update evaluations that examine whether revisions are proportionate to new evidence, whether justified commitments persist, and whether uncertainty is expressed appropriately.
OPEN QUESTION How can intermediate-state evaluations detect unsupported updates and masked failures that outcome-only scores miss?
WHERE THE MECHANISMS MEET
Following a change through the system.
A newly available tool can alter an agent's actions, create new dependencies between agents, and change where failures may spread. Those same changes affect what a monitor must record and what an evaluation must test. Following this chain requires both local traces and population-level analysis.
ANOTHER RESEARCH THREAD
Generative vision & diffusion models.
I have also worked on controllable generation and editing with diffusion and rectified-flow models, including rare and long-tail concepts, semantic disentanglement, and preference-aware personalization. Across these problems, I am interested in how language and structured guidance can make visual generation more precise and expressive.
description→Semantic
guidance→Generation
& editing