The code-agent ordering (TrackerAgent > SetRaceAgent > EconomyAgent) shows information exploitation matters more than greedy quartet-chasing, which in turn outperforms conservative budgeting

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Information exploitation matters more than greedy quartet-chasing, which in turn outperforms conservative budgeting.claim0.820
  Ordering of code agents' strengths reflects key strategic dimensions.
• Two heuristic code agents (TrackerAgent and SetRaceAgent) outperform most tested LLMs.claim0.807
  Calibration that conditional logic can beat cost-efficient LLMs in this setting.
• Code agents operate on structured data with exact arithmetic, while LLMs must parse natural-language observations and track state across turns; some failures may partly reflect numerical parsing or working-memory limitationsclaim0.763
  discussion of potential confounds
• Active inference agent with learnable preferences developed a strict preference for goals (score +) when the Frisbee location was encountered first, becoming a goal-seeking agent.finding0.759
  Figure 5.4 and text.
• Single agents can generalize better to unseen tasks because they are not constrained by predefined heuristic-based workflows.claim0.750
  Architectural belief motivating single-agent design choice; suggests flexibility provides better out-of-distribution performance.
• Within-agent score standard deviation suggests deck order matters more than seat position.claim0.748
  Observation from variance analysis, though not tested as hypothesis.
• Active inference agents can learn their own reward function (prior preferences) by interacting with the environment, bypassing the need for an explicit reward signal.claim0.746
  Abstract and §3, preference learning section.
• Smith & Boyd (1991) criteria are irrelevant across the vast majority of the space of possible agents.claim0.745
  Dismissal of earlier criteria as too narrow.