Works with the AI Agents you already use
Claude Code
Cursor
Codex CLI
Muse
Dots+21 moreClaude CodeCursorCodex CLIMuseDots
Manus
Grok
Grok Bot+18 agentsAI Agent Debugging & Repair Engineer
Diagnose and repair LLM agent failures, tool errors, RAG grounding issues, and multi-agent coordination logic.
$49.99
AI Agent Debugging & Repair Engineer
Example session with this skill installed
My LangChain agent is failing to call the 'get_user_data' tool. It just says it can't find the user, but the user ID is definitely in the DB. Here is the tool schema and the recent error log.
- Read your context and instructions
- Compiled the ai agent debugging
I have analyzed the tool failure incident. The agent is failing to pass the required user_id as an integer, instead sending it as a string, which triggers a validation error in your Pydantic model before the DB query even executes.
Incident Summary
- Symptom: Agent fails to execute
get_user_datatool; reports user not found. - Severity: High
- Impact: Blocked user profile lookups.
- Environment: LangChain v0.1.12, Python 3.10, OpenAI GPT-4-turbo.
Root-Cause Analysis
| Rank | Hypothesis | Evidence | Confidence | Confirmation test |
|---|---|---|---|---|
| 1 | Schema Type Mismatch | Logs show ValidationError for field user_id: "value is not a valid integer". | High | Call tool manually with string ID. |
| 2 | Prompt Ambiguity | System prompt doesn't specify ID format (UUID vs Int). | Medium | Check model output logs for raw JSON. |
Recommended Repair
- Minimal fix: Update the tool's JSON schema description to explicitly request an integer and add a pre-processing cast in the tool function.
- Files/components affected:
tools/user_service.py - Patch or implementation:
# Before
def get_user_data(user_id: str):
return db.fetch(user_id)
# After
def get_user_data(user_id: int):
"""Fetch user data. The user_id must be a numeric integer."""
try:
return db.fetch(int(user_id))
except ValueError:
return "Error: user_id must be an integer."
- Risks and rollback: Low risk. Rollback involves reverting schema description.
Validation Plan
| Test | Input/condition | Expected result | Pass criteria |
|---|---|---|---|
| Integer Input | user_id=101 | User data dict | 200 OK |
| String Input | user_id="101" | User data dict | Auto-cast success |
| Invalid String | user_id="abc" | Error message | Graceful failure |
Prevention
- Monitoring: Log Pydantic validation errors specifically to Sentry.
- Guardrails: Add
TypeAdaptervalidation in the tool decorator. - Regression coverage: Add a unit test to the agent test suite covering string-to-int conversion for this tool.
Next steps
- Update the Pydantic model for the
get_user_datatool to useinttype. - Add an example in the system prompt showing a successful tool call with an integer ID.
- Deploy the updated tool schema to your agent environment.
Connects securely to your tools. The creator never sees your data.
About this skill
Debugging LLM-powered agents is difficult because failures are often non-deterministic, spanning prompt ambiguity, tool schema mismatches, and RAG retrieval gaps. This skill provides a systematic engineering framework to diagnose agent failures and implement evidence-based repairs.
What it does
- Failure Classification categorize issues across prompt conflicts, tool execution, memory leaks, or RAG grounding failures.
- Root-Cause Analysis ranks hypotheses by confidence and provides specific tests to confirm or reject them.
- Repair Engineering designs minimal patches for code, prompts, or tool schemas to resolve verified causes without full rewrites.
- Validation Planning generates test cases for edge cases, malformed inputs, and tool timeouts to ensure reliability.
- Operational Safeguarding recommends specific monitoring, circuit breakers, and regression gates to prevent recurrence.
How it works
- Submit Incident Data provide the agent purpose, expected vs actual behavior, logs, and tool schemas.
- Review Triage receive a classified breakdown of the failure type, severity, and user impact.
- Analyze Root Causes review a ranked table of hypotheses backed by available evidence.
- Implement Repair apply the provided patch plan, which includes before/after behavior and rollback steps.
Frameworks & tools
Works with LangChain, LlamaIndex, CrewAI, AutoGPT, and custom LLM implementations. Compatible with OpenAI, Anthropic, and local model providers using tools, RAG pipelines, or multi-agent orchestration.
Why this beats prompting it yourself
Standard prompts often suggest speculative rewrites that break existing functionality. This skill uses a structured diagnostic workflow that separates confirmed facts from hypotheses, ensuring repairs are minimal, safe, and testable.
Use cases
- Fixing Tool Failures resolve JSON schema mismatches or argument parsing errors in agentic workflows.
- Optimizing RAG Grounding diagnose why an agent provides hallucinations despite having access to vector data.
- Debugging Loops stop agents from getting stuck in infinite tool-calling cycles or repetitive reasoning paths.
- Securing Tool Access identify and patch prompt injection vulnerabilities or unsafe tool authorizations.
Known limitations
Requires specific logs, traces, or code context to move beyond a bounded diagnosis. Cannot execute code or access live production environments directly.
Requires specific logs or traces to confirm hypotheses. Cannot access live production environments or execute code directly. Logic is bounded by the context provided in the report.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 9 days ago
Permissions
No special permissions declared or detected