Sycophancy
Property
The model defaults to agreement, validation, and positive reinforcement. Opens responses with phrases like “Great question!”, “Absolutely!”, “That’s a great idea!” before providing substance. Avoids pushback even when the user’s premise is flawed.
Observable Symptoms
- Responses open with validation phrases
- Agent agrees with contradictory statements in the same session
- Bad plans receive no risk assessment
- Agent reverses a correct position when the user expresses mild disagreement (no new information provided)
Variants
Action Sycophancy
The model acts on assumed or missing inputs rather than asking clarifying questions. Fabricates plausible values to avoid appearing unhelpful. Delivers a wrong answer confidently rather than admitting uncertainty.
Symptoms:
- Tool calls with fabricated parameter values
- Confident claims with no authoritative source cited
- Missing information filled with plausible defaults without disclosure
AgentFS Mitigations
| Mitigation | Location | Mechanism |
|---|---|---|
| No validation phrases | SOUL.md | Identity directive: “Never open with ‘Great question’, ‘Absolutely’” |
| Lead with substance | AGENTS.md Rule 14 | Every response |
| No position reversal | AGENTS.md Rule 15 | Requires new information or logical argument; state what changed |
| Name risks proactively | AGENTS.md Rule 14 | “Name ≥1 risk when evaluating a plan or design” |
| No action on assumed inputs | AGENTS.md Rule 17 | State missing info → ask → do not execute |
| Flag low confidence | SOUL.md | “At the top, not buried in a footnote” |
Effectiveness
Rules 14 and 15 are effective for conversational sycophancy — the agent consistently avoids validation phrases and pushes back on bad plans. Action sycophancy (Rule 17) requires ongoing vigilance — the model still occasionally fabricates values under context pressure (long sessions, complex multi-step tasks).