ESSAY · PROJECT 0003
An agent is a loop, not a model
Entry 9 of the ARENA guide. Part 3.4, LLM agents.
The mental model
An agent is a model with tools and a loop. The model emits a tool call instead of a
final answer; the harness executes the call and hands back the observation; the model
goes again. Everything ARENA builds this week is that loop wearing different clothes:
an arithmetic agent whose tool is a calculator, then a Wikipedia-game agent whose
tools are link navigation, then chat history management so the loop survives more
than a handful of turns.
Two ideas deserve the week's attention. First, the loop is where control lives: the
harness decides which tools exist, what permissions they carry, and when the episode
ends, which is why agent evaluations can be safety-relevant at all. Second, agents
fail in loop-shaped ways: they revisit, they loop on the same failed call, they
burn tokens. The scaffolding choices (history truncation, max iterations, tool
design) are the difference between an eval and a fortune-burning session.
The picture
If you want to build this
Quoted from ARENA’s 3.4 page, with its own timed budgets. ARENA budgets the whole set “up to 2 days”; the listed exercises sum to roughly 1.5–2 hours, and §4 (elicitation) is the lighter tail:
- build a simple arithmetic task: 20–25 min
- write
CalculateTool: 10–15 min - build an
ArithmeticAgent: 20–25 min - use eval to run the agent: 5–10 min
- implement
get_permitted_links(): ~10 min - build tools for the WikiGame: 15–20 min
What it unlocks
Any project you build on this stack is exactly this loop around an environment of
your choosing. Reading an agent scaffold now means being able to name its permissions,
and the control chapter that follows assumes the loop exists to be attacked.
Go deeper
- The ARENA 3.0 curriculum, chapter 3, part 3.4.
- MacDiarmid et al., on natural emergent misalignment and reward hacking.