Which two terms in Russell and Norvig's definition separate an agent from a single model call?
An agent is a perceive–decide–act loop, and the guardrails sit on every turn.
Stuart Russell, Peter Norvig · Artificial Intelligence: A Modern Approach
Why is Simon's ant an argument for context engineering?
The environment the agent sees explains more of its behaviour than the model does.
Herbert A. Simon · The Sciences of the Artificial
In Sutton's terms, what bet does an agent with many hand-written rules make?
An agent's scaffold must be thin enough that a better model improves it rather than breaks it.
Richard S. Sutton · The Bitter Lesson
What is the difference between a workflow and an agent, as defined in "Building effective agents"?
If you can draw the flowchart, build a workflow, not an agent.
Erik Schluntz, Barry Zhang · Building effective agents
What does ReAct add over plain chain-of-thought?
Thought, action, observation: every turn of the loop is a place to check and a place to attack.
Shunyu Yao et al. · ReAct: Synergizing Reasoning and Acting in Language Models
When is it worth splitting one agent into several?
The number of agents follows the number of distinct contexts, not the org chart.
Marvin Minsky · The Society of Mind
What does MCP standardise, and who has governed it since December 2025?
The tool description is the prompt you did not write, yet the model reads it on every turn.
Jon Postel · RFC 793 — Transmission Control Protocol
What is the difference in role between MCP and A2A?
The boundaries between agents will be the boundaries between teams, unless you choose them.
Melvin E. Conway · How Do Committees Invent?
What do you ask of an agent framework before adopting it?
Start from the hundred-line raw loop; adopt a framework when you need durable state, human pauses or routing.
John Gall · Systemantics: How Systems Work and Especially How They Fail
What is "tool poisoning" and why does the user not see it?
A tool's description is code executed by the model: review it, pin it, isolate it.
Ken Thompson · Reflections on Trusting Trust
What is the recommended order of cost-reduction measures in an agent?
The expensive model is called on exception, not by default.
Alfred North Whitehead · An Introduction to Mathematics
What procedure does the NVIDIA paper propose for moving an agent from large models to small ones?
The planner can be large; the workers can be small, specialised and local.
Peter Belcak et al. (NVIDIA Research) · Small Language Models are the Future of Agentic AI
What does LoRA freeze and what does it train?
LoRA freezes the model and learns a small low-rank correction, so an adapter is a file, not a model.
Edward J. Hu et al. · LoRA: Low-Rank Adaptation of Large Language Models
Name three cases where LoRA beats RAG and prompting.
Fine-tuning for behaviour, RAG for knowledge, a prompt for everything it can carry.
John Schulman et al. · LoRA Without Regret
What does Goodhart's law mean for an LLM evaluator?
An LLM judge is calibrated on humans and kept out of the optimisation loop.
Marilyn Strathern · «Improving ratings»: audit in the British University system
What are the two solid arguments for on-prem inference?
On-prem is justified by sustained utilisation or by data that cannot leave, not by the token price.
William Stanley Jevons · The Coal Question
What are the three features of the lethal trifecta?
Break one leg of the trifecta: no outbound channel, no untrusted content or no private data.
Simon Willison · The lethal trifecta for AI agents
Why does the agent not get the user's full token?
An identity of its own for the agent, allowlisted rights, a sandbox and approval on the irreversible.
Jerome H. Saltzer, Michael D. Schroeder · The Protection of Information in Computer Systems
Where does personal data enter an agent, and where do the measures go?
Pseudonymise at the tool boundary, re-identify at the end, keep the traces clean and the memory off.
Parlamentul European și Consiliul · Regulamentul (UE) 2016/679 — GDPR
By what criterion is an agent's level of autonomy set?
Autonomy is granted by reversibility: suggest, ask for approval, act and report.
Joseph Weizenbaum · Computer Power and Human Reason