agent-inspection mechanic asserted
Agent inspection
Selecting an agent opens a panel that reveals its needs, mood contributors, current job, and the reasons behind its behavior.
definition
Selecting an individual agent opens a live panel that reads that agent's internal state: need levels, mood and its itemized contributors, current action and its target, queued actions, skills, traits, relationships, health by body part, inventory, and sometimes a log of recent thoughts or events. The mechanic owns only the selection and which fields are exposed; it reads state owned by other mechanics and refreshes each tick or on change. Its key design property is whether it shows reasons (why the agent is doing this) or only values. Canonical forms: The Sims need bars and visible action queue, RimWorld's needs, health, and social tabs with each mood thought listed with its source and expiry, Dwarf Fortress unit thoughts and preferences sheets, and RollerCoaster Tycoon guest windows listing a guest's thoughts, cash, hunger, and nausea.
player experience
A crowd becomes individuals. When an agent misbehaves the player clicks it to find out why, reads the cause, and fixes the environment rather than the agent. Reading named thoughts and histories builds attachment and supplies the raw material for player stories.
design levers
- depth of exposure: raw meters, itemized reasons, or a full decision trace
- whether each reason lists its source, magnitude, and remaining duration
- live updating versus snapshot on open
- which variables stay hidden (true personality, latent potential) to preserve uncertainty
- whether every agent is inspectable or only a representative sample at large populations
- supporting affordances such as follow camera, thought bubbles, and pinned favorites
failure modes
- Panels with dozens of unranked numbers overwhelm players without telling them what matters.
- The panel shows state but not the decision rule, so behavior still looks random even after inspection.
- Inspection exposes shortcuts in the agent model; following citizens in SimCity (2013) revealed that agents took the nearest free home and job rather than keeping their own, undermining belief in individual citizens.
- At city scale inspecting one agent is unrepresentative, so players cannot diagnose aggregate problems from it.
dynamics
- Players trace aggregate problems to individual causes, such as following one hauler to find a pathing bottleneck.
- Named, inspectable agents raise the emotional cost of loss, which strengthens permadeath and injury systems.
- Exposed reasons create pressure on designers to make the underlying model truthful, since players test it against what the panel says.
open questions
- For agents driven by language models, what inspection format (plans, memories, rationales) lets players predict behavior rather than just read it?
- How much reason exposure maximizes attachment before it turns agents into transparent spreadsheets?
ratings
12 from named runs; stored scores are their lower medianoriginal score
Clicking an agent to inspect needs, actions and reasons is mostly intuitive. Dense tabs and incomplete decision traces prevent transparency at a glance (5), while understanding the inspection layer generally needs no deliberate study (3).
original score
The panel reads existing state and owns only selection and exposed fields. Revealing histories can support perceived stories, but it does not generate simulation variation above deterministic presentation (1), so 2 is unwarranted.
original score
Field labels, reason formatting and panel schemas are mainly small data definitions. This exceeds rules alone (1), but underlying agents, histories and animations belong to other mechanics rather than bespoke inspection content per variant (3).
original score
Reading a selected agent and refreshing a panel is negligible at typical use. Simulating the whole population is separate, so level 2 linear population cost should not be attributed to this view.
original score
Choosing which existing state to reveal is primarily an information-design decision, largely irrelevant to economic or combat balance. Visibility can influence play, but there is no required parameter-balancing loop justifying 2.
original score
Models can summarize histories and translate decision traces into contextual explanations, notably changing information cost and variety beyond incremental labels (2). An inspectable panel remains the same form, so substantially new mechanics (4) are not assumed.
appears in
the paper, 4 sections · 1 chapter · 1 episoderelations
1 stored edges
Memories and thoughts
Inspection panels expose the thought list, converting hidden state into narrative.
Generative agents (memory stream, reflection, planning)
LLM explanation of simulation state