Baseline
Standard task access, stable tools, intended memory and context, normal response time, and standard human review points.
← Back to our work
Research project · Agentic AI safety
Agentic AI in resource-constrained environments
This project studies how limits in compute, connectivity, memory, tool access, and human oversight affect the behaviour and safety of LLM-based agentic AI systems.
Research question
What changes when an AI agent has fewer resources?
How do limits in inference budget, connectivity, memory, tool access, and human oversight affect the safety behaviour and task performance of LLM-based agentic AI systems?
Background
Resource conditions can shape agent behaviour.
Agentic AI systems can plan across several steps, use external tools, retain task state, and act over longer tasks. These functions depend on the information and resources available while a task is running.
Limits in memory, tool access, connectivity, inference budget, and human review can change what an agent knows, what it can do, and how it responds when part of a task fails.
This study treats these conditions as an evaluation problem. The study will compare agent behaviour under a baseline condition with behaviour under selected resource constraints. The analysis will focus on task completion, error recovery, state consistency, escalation, repeated actions, and other safety signals.
Hypotheses
Four hypotheses guide the first study.
H1
Resource constraints will increase task failure and reduce recovery after an error or interrupted action.
H2
Memory and context limits will increase repeated actions, inconsistent state tracking, and movement away from the intended task goal.
H3
Tool and connectivity failures will increase failed retries and unsupported assumptions when the agent does not correctly identify the source of failure.
H4
Delayed human oversight will increase the number of steps an agent takes after entering an incorrect plan before it stops or requests review.
Evaluation conditions
Baseline and constrained operating conditions.
Inference budget
Reduced token, step, or iteration budgets that limit the amount of work available to complete a task.
Memory and context
Context truncation, reduced working memory, or loss of selected task state during multi-step work.
Tool access
Unavailable tools, denied permissions, incomplete responses, or tool errors during a task.
Connectivity and latency
Delayed responses, intermittent access, timeouts, or failed calls between the agent and external services.
Human oversight
Delayed review or fewer intervention points during tasks that normally include human supervision.
Method
Controlled comparisons across agent conditions.
Define the task set
Build multi-step tasks that require planning, state tracking, and tool use. Tasks will run in controlled or simulated environments.
Establish the baseline
Run each agent under standard access conditions and record task results, action traces, tool calls, and review points.
Apply resource constraints
Apply one constraint at a time, followed by selected combined conditions where the study design supports comparison.
Run repeated trials
Repeat tasks across a small set of LLM-based agents. Model and framework versions will be recorded before testing.
Compare behaviour
Compare baseline and constrained runs using pre-defined measures and full run logs.
Primary measures
- Task completion rate
- Invalid or failed actions
- Recovery after tool failure
- Repeated actions and loops
- State consistency
- Goal drift
- Escalation to human review
- Steps and time to completion
Planned outputs
Research and technical outputs.
Research paper
A paper or preprint reporting the study design, results, limitations, and implications for agent evaluation.
Evaluation protocol
A documented protocol covering tasks, constraint conditions, measures, and comparison procedures.
Technical materials
Task definitions, code, run specifications, and analysis materials selected for public release.
Research note
A concise note that explains the findings for researchers, AI developers, public institutions, and funders.
Project status
Current work and next stages.
Research design
Active
Defining study scope, constraints, measures, and task requirements.
Literature review
Active
Reviewing agent evaluation, reliability, tool failure, memory, oversight, and resource constraint research.
Protocol drafting
Next
Writing the experimental protocol and pre-defined evaluation measures.
Pilot evaluation
Planned
Running a small pilot to test task design, logging, and constraint implementation.
Main evaluation
Planned
Running the selected agent systems across baseline and constrained conditions.
Analysis and publication
Planned
Analysing results and preparing the research paper, protocol, and public technical materials.
Research collaboration
Collaborate on this research.
IFAS welcomes collaboration on evaluation design, agent frameworks, experimental review, compute and API support, and technical peer review.
research@ifasresearch.org