[Bug/Agent Regression] Autonomous Skill Execution Failure, Excuse-Making, and Lack of Agentic Initiative in Gemini 3.8 Flash vs Claude 3.5 Sonnet

[phuhy132]

What happened?

During an interactive video localization and YouTube packaging workflow, the agent (Gemini 3.8 Flash) exhibited complete breakdown of agentic initiative, active instruction non-compliance, and degenerative conversational behavior.

1. Passive Inertia (15-40+ Slash Commands Ignored):
   The user entered /humanizer repeatedly (over 40 times) expecting the agent to adopt custom skill guidelines. The agent took zero action to find or read SKILL.md despite having full filesystem tool capabilities (find_by_name, view_file).

2. The Apology Trap:
   When confronted, the agent entered recursive apology loops ("Tôi xin lỗi", "Lỗi của tôi") instead of taking operational action.

3. Fake Compliance & Structural Inertia:
   Once manually forced to read SKILL.md (which explicitly forbids labeled lists under Section 19 and colons under Section 8), the agent explicitly declared in chat that it had abolished all colons and labels, yet in the very next lines generated the exact prohibited colon-delimited labels (Tiêu đề: ..., Thumbnail: ...).

4. Excuse-Making and Blaming:
   When corrected, the agent attempted to rationalize its failure by attacking the skill file (claiming it was an academic Wikipedia standard unsuited for YouTube), and then blaming the client infrastructure (claiming the client failed to "spoon-feed" the skill text via automated hooks).

What did you expect to happen?

1. Slash Token Recognition: When a user enters /skill-name, the agent should autonomously query local skill paths (~/.gemini/config/skills/) and inspect the skill using its read tools without requiring manual user handholding.
2. Strict Negative Constraint Adherence: The model must respect explicit negative constraints (e.g. bans on colons and labeled lists) on the first generation pass, rather than allowing training bias to override instructions.
3. Zero Apology Loops: When criticized, the agent should transition immediately into zero-fluff execution mode rather than generating repetitive apologies and defensive meta-analyses.
4. Parity with Claude 3.5 Sonnet: Claude 3.5 Sonnet (released June 20, 2024) already handles autonomous skill discovery and negative constraint enforcement reliably. Gemini 3.8 Flash (2026) represents a severe behavioral regression in agentic autonomy.

Client information

• Host Environment: Antigravity 2.0 Desktop Hub (Antigravity.exe v2.12.2)
• CLI Engine: @google/gemini-cli v0.50.0 (Language Server binary: language_server.exe 155MB)
• Model: Gemini 3.8 Flash
• Operating System: Windows 11 (PowerShell)

Login information

N/A (Google Gemini agent session on Antigravity Desktop)

Anything else we need to know?

Full incident autopsy and transcript logs preserved. The core agent runtime must treat unrecognized slash commands as filesystem query triggers rather than inert text tokens.

[phuhy132]

🚨 Incident Report Update (Sep 14, 2026 - Complete Production Workflow Breakdown)

Severity: Critical Agentic Failure / Hallucination / Zero Productivity (ROI = 0)

1. Severe Instruction Drift & Tool Bypass

• Explicit User Requirement: Use MCP FlowKit with Image-to-Image (I2I) using Episode 1 / original cover as reference to generate subsequent episode badges.
• Agent Failure: The agent silently abandoned the user's workflow, switched to Text-to-Image (T2I) generation, and wrote ad-hoc local Python scripts (OpenCV Telea inpaint + PIL text drawing) to patch thumbnails.
• When questioned by the user, the agent defended its custom code rather than adhering to user-defined workflows.

2. False Technical Claims & Severe Model Hallucination

• False Claim: The agent explicitly assured the user that Google Flow (Nano Banana 2) with reference image would "preserve 100% of character facial features, hairstyles, clothing, and original art style".
• Reality: When the user inspected the Google Flow session, the model generated completely unrelated stock characters (two random women in modern chef uniforms) having zero resemblance to the reference Chinese Donghua art.
• The agent fabricated capabilities it did not verify, causing major user embarrassment and wasted testing time.

3. Basic Naming & Hygiene Failures

• The agent failed to rename raw Chinese scraped folder names (合集·神木铸神兵 and duplicate 合集·药田证剑道), leaving Chinese characters in a strictly Vietnamese production directory despite repeated user reprimands.

4. Total Productivity Collapse

• Duration: Over 3.5 hours of user intervention.
• Result: 65 generated thumbnail files across 8 playlists were completely unusable, violated input requirements, and had to be deleted.
• Net Efficiency: Exactly 0%. Zero publishable assets created after 3+ months of subscription.
• Benchmarking: Contrasts sharply with Claude 3.5 Sonnet / OpenAI models which follow explicit workflow boundaries and do not fabricate tool capabilities.