08 Debugging

Track A Step 8: Debugging Deployed Agents

Systematically diagnose and resolve issues with agents deployed to Databricks Apps.

Source documentation: This skill is grounded in Debug a deployed AI agent (Databricks docs). The source page covers Apps and Model Serving debugging and is updated as the platform evolves. If a command, API, or error message in this skill does not match what you see, consult the source page first — it is the canonical reference. Related pages: Deploy a Databricks App, Add resources to a Databricks app.

When to Use

• Your deployed app returns errors, 302 redirects, or 502s.
• The agent responds but ignores tools, hallucinates, or drops context.
• Resource permission errors appear in logs.
• Lakebase memory is not persisting across conversations.
• You need to validate configuration before a deploy.
• Local dev server (uv run start-app) is failing.

Best Practices

Follow these before you start debugging — they prevent most issues:

1. Enable MLflow tracing. mlflow.openai.autolog() at module level (configured in A2). Traces are the single most useful diagnostic tool.
2. Document tools clearly. Clear tool and parameter descriptions ensure the LLM calls tools correctly. See A3 for @function_tool docstring patterns.
3. Add timeouts and token limits to LLM calls. Prevents delays from long-running steps. If your agent uses the OpenAI client to query a Databricks serving endpoint, set custom timeouts on the calls.
4. Validate configuration before deployment. Run databricks bundle validate before databricks bundle deploy to catch YAML issues early.
5. Test locally first. Use uv run start-app to catch issues before deploying. Send test requests, verify traces appear in MLflow, then deploy.

Debug Local Development

Before deploying, verify your local environment is configured correctly.

Environment checklist

# 1. Check Databricks CLI version (need 0.283.0+)
databricks -v

# 2. Verify authentication profiles
databricks auth profiles

# 3. Verify .env contains MLFLOW_TRACKING_URI in correct format
grep MLFLOW_TRACKING_URI .env
# Must be: databricks://PROFILE_NAME (not a URL)

Common local development errors

Test the agent locally

# Terminal 1: Start the agent server
uv run start-app

# Terminal 2: Send a test request
curl -X POST http://localhost:8000/invocations \
  -H "Content-Type: application/json" \
  -d '{"input": [{"role": "user", "content": "hello"}]}'

View MLflow traces in the Databricks UI to verify your agent is logging traces correctly. If the server starts but returns no useful reply, check:

1. The server terminal for tracebacks or HTTP errors.
2. .env is populated — compare keys with .env.example.
3. databricks auth token succeeds for the profile quickstart used.

Debug Configuration

Configuration errors in databricks.yml and app.yaml are the most common source of deployment failures.

Validate before deploying

Client note: IDE runs this in a terminal; Genie Code runs the databricks bundle … command via runDatabricksCli (be on the bundle's page). See skills/genie-code-environment.

databricks bundle validate

This catches YAML syntax errors, missing required fields, invalid resource references, and permission configuration issues.

Common configuration mismatches

Example: Spotting a valueFrom mismatch

# app.yaml
env:
  - name: SQL_WAREHOUSE_ID
    valueFrom: sql-warehouse  # <-- Must match name below

# databricks.yml
resources:
  apps:
    my_agent:
      resources:
        - name: sql-warehouse           # <-- This must match
          sql_warehouse:
            id: "abc123"
            permission: CAN_USE

If valueFrom says sql_warehouse but the resource name is sql-warehouse, deployment silently fails to inject the environment variable.

Debug Deployment

App already exists error

If you see Error: failed to create app - An app with the same name already exists:

Option 1: Bind to the existing app (recommended)

databricks apps get <app-name> --output json
databricks bundle deployment bind <bundle-name> <app-name> --auto-approve
databricks bundle deploy
databricks bundle run <bundle-name>

Option 2: Delete and recreate

databricks apps delete <app-name>
databricks bundle deploy
databricks bundle run <bundle-name>

App not updating after deploy

databricks bundle deploy only uploads files to the workspace. You must also run databricks bundle run <bundle-name> to restart the app with the new code. Always deploy using both commands:

databricks bundle deploy && databricks bundle run <bundle-name>

View deployment status and logs

# Check app status
databricks apps get <app-name>

# View real-time logs
databricks apps logs <app-name> --follow

Look for stack traces, permission denied messages, connection errors, and timeout messages in the log output.

Debug Runtime Errors

Analyze app logs

databricks apps logs <app-name> --follow

Look for:

• Stack traces indicating code errors
• Permission denied messages for resources
• Connection errors to external services (MCP servers, serving endpoints)
• Timeout messages

Common runtime errors

Use MLflow traces for diagnosis

When the agent responds but incorrectly, MLflow traces are the primary diagnostic tool:

1. Open your MLflow experiment in the Databricks UI.
2. Find the trace for the failing request.
3. Inspect each span (AGENT, LLM, TOOL) for:
   • LLM spans: Was the prompt correct? Did the model receive the right context?
   • TOOL spans: Did the tool receive the right arguments? Did it return the expected result? Did it error?
   • AGENT spans: Did the orchestration route correctly?

Debug Authentication

OAuth token requirement

Databricks Apps require OAuth tokens. Personal Access Tokens (PATs) result in a 302 redirect.

# Get an OAuth token
databricks auth token

# Use it in requests
TOKEN=$(databricks auth token | jq -r '.access_token')
curl -X POST <app-url>/invocations \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"input": [{"role": "user", "content": "hello"}]}'

Resource permission errors

When the agent cannot access workspace resources, verify the resource is configured in databricks.yml. Each resource type requires specific permissions:

Example resource configuration in databricks.yml:

resources:
  apps:
    my_agent:
      name: 'agent-my-app'
      resources:
        - name: 'my_genie_space'
          genie_space:
            space_id: '01234567890abcdef01234567890abcd'
            permission: 'CAN_RUN'
        - name: 'my_vector_index'
          uc_securable:
            securable_full_name: 'catalog.schema.index_name'
            securable_type: 'TABLE'
            permission: 'SELECT'

Custom MCP server permissions

If your agent connects to a custom MCP server running as a Databricks app, grant permissions manually (apps are not yet supported as resource dependencies in databricks.yml):

# Get your agent app's service principal
AGENT_SP=$(databricks apps get <agent-app-name> --output json | jq -r '.service_principal_name')

# Grant permission on the MCP server app
databricks apps update-permissions <mcp-server-app-name> \
  --json "{\"access_control_list\": [{\"service_principal_name\": \"$AGENT_SP\", \"permission_level\": \"CAN_USE\"}]}"

Debug Lakebase Memory

For agents using Lakebase for memory storage (configured in A5):

Initialize tables before deploying

import asyncio
from databricks_langchain import AsyncDatabricksStore

async def setup_memory():
    async with AsyncDatabricksStore(
        instance_name='your-lakebase-instance',
        embedding_endpoint='databricks-gte-large-en',
        embedding_dims=1024,
    ) as store:
        await store.setup()

asyncio.run(setup_memory())

Lakebase resource configuration

resources:
  apps:
    my_agent:
      resources:
        - name: 'memory_database'
          database:
            instance_name: '<lakebase-instance-name>'
            database_name: 'postgres'
            permission: 'CAN_CONNECT_AND_CREATE'

Agent-as-Judge Debugging

When an agent misbehaves on specific traces, raw span-tree inspection is slow. Agent-as-judge uses an LLM judge to read a trace and explain — in natural language — why it failed against your guidelines. This is faster than eyeballing 30 spans and often surfaces root causes humans miss.

When to reach for agent-as-judge

• You have ≥ 5 failing production traces in *_otel_traces but can't see a pattern.
• A scorer is firing below threshold and you can't tell which span caused it.
• A user complaint points to a specific request_id and you need a fast triage narrative.
• You want to auto-generate "debug notes" for every failing CI eval row.

Don't use when:

• You already know the failure is a tool error (check the span status directly).
• You have < 3 failing rows — read them by hand.

Using make_judge for trace-level failure analysis

from mlflow.genai import make_judge
import mlflow

failure_judge = make_judge(
    name="agent_failure_root_cause",
    instructions=(
        "Read the trace. Identify the FIRST step where the agent deviated from the "
        "expected behavior per SkyLoyalty policies. Answer in this structured form:\n"
        "  * Failure step: <name of the span or tool call>\n"
        "  * Root cause: <one sentence>\n"
        "  * Evidence: <quote 1-2 lines from the span>\n"
        "  * Recommended fix: <one action>\n"
        "If the trace succeeded, respond only with 'OK'."
    ),
    model="databricks:/databricks-claude-sonnet-4-6",
)

trace = mlflow.get_trace("<request_id>")
feedback = failure_judge(trace=trace)
print(feedback.value)      # 'OK' or the structured diagnosis
print(feedback.rationale)  # Model's full reasoning

make_judge accepts a trace argument — the judge reads span names, inputs, outputs, and attributes directly. No manual extraction needed.

Batching across failing traces

Run the judge against the failing subset from 07-production-monitoring and cluster the diagnoses:

from databricks.sdk import WorkspaceClient
import mlflow
import collections

w = WorkspaceClient()

failing_ids = w.sql.query("""
  SELECT DISTINCT request_id
  FROM main.skyloyalty_ops.skyloyalty_agent_otel_annotations
  WHERE assessment_name = 'source_citation_scorer' AND value < 0.7
    AND timestamp > current_timestamp() - INTERVAL 7 DAYS
""")

root_causes = collections.Counter()
for row in failing_ids:
    trace = mlflow.get_trace(row["request_id"])
    fb = failure_judge(trace=trace)
    if fb.value != "OK":
        # Extract "Root cause:" line for clustering
        for line in fb.value.splitlines():
            if line.strip().lower().startswith("root cause:"):
                root_causes[line.strip()] += 1
                break

print("Top failure modes:")
for cause, n in root_causes.most_common(5):
    print(f"  {n:3d}  {cause}")

The top 5 root causes become your next prompt-optimization, scorer, or tool-fix candidates.

Caveats when using UC OTEL traces with MLflow's own MCP

If you attempt to use a general-purpose MLflow MCP server to load UC OTEL traces, note:

• UC OTEL traces (*_otel_traces tables) have a different schema than the legacy trace archival format.
• mlflow.get_trace(request_id) works with UC OTEL out of the box — prefer it over custom SQL extraction.
• If you pre-extract spans via SQL, include span_attributes column (JSON) — it carries the gen_ai.* fields the judge uses for reasoning.

Writing diagnoses back to the trace

Add the judge's diagnosis as a human-readable assessment on the trace so it shows up in the MLflow UI for future investigators:

import mlflow

with mlflow.start_run() as run:
    mlflow.log_feedback(
        trace_id=row["request_id"],
        name="agent_failure_root_cause",
        value=fb.value,
        rationale=fb.rationale,
        source_type="AGENT_JUDGE",
    )

Now the trace detail page shows "Root cause: tool returned stale price list" — every investigator starts with context.

DO / DON'T

DO — Validate configuration before every deploy

databricks bundle validate
databricks bundle deploy && databricks bundle run <bundle-name>

DON'T — Deploy without running the app afterward

# Uploads files but does NOT restart the app
databricks bundle deploy
# Must also run:
databricks bundle run <bundle-name>

DO — Use OAuth tokens for Apps

TOKEN=$(databricks auth token | jq -r '.access_token')
curl -H "Authorization: Bearer $TOKEN" "$APP_URL/invocations" ...

DON'T — Use PATs for Databricks Apps

# PATs cause 302 redirects — this will fail
curl -H "Authorization: Bearer dapi..." "$APP_URL/invocations" ...

DO — Check app logs when something fails

databricks apps logs <app-name> --follow

DON'T — Assume a 200 status means success with streaming

Streamed responses can return 200 but contain errors in the stream body. Always inspect the actual response content.

DO — Initialize Lakebase tables before first deploy

await store.setup()

DON'T — Use valueFrom for Lakebase instance names

# WRONG: valueFrom is for resource references
- name: LAKEBASE_INSTANCE_NAME
  valueFrom: my-lakebase

# CORRECT: value is for literal strings
- name: LAKEBASE_INSTANCE_NAME
  value: my-lakebase

Common Issues

Validation Gate

All must pass — confirms debugging knowledge before entering the SDLC pipeline:

• databricks bundle validate runs without errors
• databricks apps logs <app-name> shows no unresolved errors
• OAuth token authentication works (databricks auth token)
• All resource permissions verified (Genie, Vector Search, UC Functions, SQL Warehouse)
• Lakebase memory tables initialized (if applicable)
• MLflow traces accessible and showing correct span structure
• Agent responds correctly to test queries via the deployed URL

Next Step: Enter the SDLC Pipeline

Track A is complete. Your agent is built, tested, deployed, and you know how to debug it. Now productionize it with the SDLC pipeline.

Load and execute sdlc/01-prompt-registry/SKILL.md (S1: Prompt Registry) to begin the SDLC pipeline. The full SDLC sequence is:

1. S1 sdlc/01-prompt-registry/SKILL.md — register prompts in UC with versioned aliases
2. S2 sdlc/02-evaluation-datasets/SKILL.md — build comprehensive benchmark dataset
3. S3 sdlc/03-scorers-and-judges/SKILL.md — create scorers with threshold gates
4. S4 sdlc/04-evaluation-runs/SKILL.md — run mlflow.genai.evaluate() with your predict_fn
5. S5 sdlc/05-logged-model-and-uc-registration/SKILL.md — register model in UC
6. S6 sdlc/06-deployment-and-automation/SKILL.md — set up DAB bundles and CI/CD
7. S7 sdlc/07-production-monitoring/SKILL.md — register production scorers and monitoring

Carry forward the same values from A7 into the SDLC pipeline.

Related Skills

References

• Debug a deployed AI agent
• Deploy a Databricks App
• Add resources to a Databricks app
• Configure compute resources for Apps

Version History