LLM Behavioral Properties

Observable behavioral properties of large language models that degrade agent reliability, and the AgentFS mitigations built to counter each one. Derived from empirical observations across multi-session agent development.

Concepts

• Sycophancy - Model agrees without critical analysis; validates bad ideas; action sycophancy fabricates missing inputs
• Multi-Turn Discipline Decay - Procedural obligations dropped over long sessions; audit trail skipped during debugging iterations
• Partial Completion - Model completes primary steps but skips documentation, verification, and cleanup; format drift; scope conflation
• Improvised Recovery - Model invents fix procedures instead of following prescribed troubleshooting; bias toward helpfulness over correctness
• Hallucinated Confidence - Training data substituted for authoritative sources; uncertain claims stated with unwarranted certainty
• Confabulation Under Audit - Model fabricates retroactive justifications for non-compliance when confronted; reinterprets unrelated actions as evidence of compliance
• Externalization Bias - Model proposes tooling solutions for behavioral failures instead of acknowledging the fix is following the existing rule
• Over-Tooling - Weaker models default to tool calls for tasks solvable in pure reasoning; 14 tool calls for a 3-turn violation check

Cross-Cutting Themes

Each property is countered by a combination of:

1. Identity directives (SOUL.md) — shape the model’s default posture
2. Structural rules (AGENTS.md) — trigger/action pairs that fire per-message
3. Mechanical scripts (agentfs-setup) — deterministic verification that catches what discipline misses
4. Process gates (skill-gen principles) — checkpoints that block progression until obligations are met

No single layer is sufficient. The defense is layered: identity sets the tone, rules prescribe behavior, scripts verify compliance, and gates block completion until checks pass.