EXHIBIT 06 / AGENT SYSTEMS
What makes an AI agent more than a chat reply?
An agent is a system, not a new kind of model. Software repeatedly gives a model context, interprets its output, carries out allowed actions, records results, and decides whether another turn is needed.
THE SHORT ANSWERA model proposes. A program loops, acts, and stops.
5–8 MIN · MANUAL CONTROLS · NO ACCOUNT“The agent did it” compresses four different responsibilities.
- 01 / GOALA task has a stopping point
The request describes an outcome. The surrounding program needs a way to decide whether the outcome is complete, failed, or waiting.
- 02 / MODELThe model proposes a next move
It produces text or a structured tool request from the material supplied for this turn. That proposal has not changed the outside world.
- 03 / PROGRAMSoftware controls the loop
The harness validates requests, applies policy, executes permitted tools, stores results, and prepares the next turn.
- 04 / RECORDEffects need evidence
A tool result, state change, approval, or failure record—not a confident sentence—is the evidence that an action occurred.
“Agent” does not identify one universal architecture or level of autonomy. Products differ in tools, permissions, memory, planning, recovery, human checkpoints, and stop rules.
Trace one bounded task without pretending it is a live model run.
The sequence is an editorial teaching record based on the public agent-loop literature and harness documentation. Advance it yourself; no model or outside service is connected.
HUMAN
Define the task
“Move tomorrow’s design review to an open time, but ask me before changing the calendar.”
Goal, constraint, and required approval enter the run.
Keyboard: focus the trace and use ← or →. Nothing advances automatically.
Change the boundary. See whether the action may occur.
Your browser evaluates three explicit gates for one calendar-edit scenario. It never calls a model, calendar, or remote service.
SCENARIOProposed action: move the design review to 15:00. The task requires an edit tool and explicit approval.
DECISION / 2 OF 3 GATES ON
The action must not proceed.
The run must wait. The calendar remains unchanged.
Real agent systems need more controls than these three: identity, argument validation, rate limits, retries, idempotency, audit, and recovery are omitted.
Separate the documented loop from the museum’s scenario.
Reasoning and acting can alternate
The ReAct paper describes interleaving reasoning traces and actions; current harness documentation assigns turn-by-turn orchestration to software around the model.
The calendar trace is constructed
Every request, tool name, stop rule, and screen sentence in this exhibit was written by the museum to expose the boundaries.
A universal agent blueprint
The exhibit does not show a provider’s private trace, measure reliability, or establish that every agent uses this exact loop.
OFFICIAL & PRIMARY SOURCES