entityprocess/agentv-trace-analyst

AgentV Trace Analyst

Analyze evaluation traces headlessly using agentv inspect primitives and jq.

Primitives

# List result files (most recent first)
agentv inspect list [--limit N] [--format json|table]

# Show results with trace details
agentv inspect show <result-file> [--test-id <id>] [--tree] [--format json|table]

# Percentile statistics
agentv inspect stats <result-file> [--group-by target|suite|test-id] [--format json|table]

# A/B comparison between runs
agentv compare <baseline.jsonl> <candidate.jsonl> [--threshold 0.1] [--format json|table]

Analysis Workflow

1. Discover results

agentv inspect list

Pick the result file to analyze. Most recent is first.

2. Get overview

agentv inspect stats <result-file>

Read the percentile table. Key signals:

• score p50 < 0.8: Significant quality issues
• latency p90 > 30s: Performance bottleneck
• cost p99 spike: Outlier cost tests to investigate
• tool_calls p90 >> p50: Some tests are much chattier

3. Investigate failures

agentv inspect show <result-file> --format json | jq '[.[] | select(.score < 0.8) | {test_id, score, assertions: [.assertions[] | select(.passed | not)], trace: {tools: (.trace.tool_calls | keys)}, duration_ms, cost_usd}]'

For each failing test, examine:

• assertions (failed): What criteria were not met? (filter for passed: false)
• trace.tool_calls: Did the agent use expected tools?
• duration_ms: Did it time out or run too long?
• reasoning: Why did the grader score it low?

4. Inspect specific tests

# Flat view with trace summary
agentv inspect show <result-file> --test-id <id>

# Tree view (if output messages available)
agentv inspect show <result-file> --test-id <id> --tree

The tree view shows the agent's execution path — LLM calls interspersed with tool invocations. Look for:

• Excessive tool calls: Agent looping or exploring unnecessarily
• Missing tools: Expected tool not called
• Long durations: Specific tool calls that are slow

5. Compare runs

agentv compare <baseline.jsonl> <candidate.jsonl>

Look for:

• Wins vs losses: Net improvement or regression?
• Mean delta: Overall direction of change
• Per-test deltas: Which tests regressed?

6. Group analysis

# By target provider
agentv inspect stats <result-file> --group-by target

# By suite
agentv inspect stats <result-file> --group-by suite

Compare providers side-by-side: which is cheaper, faster, more accurate?

Advanced Queries with jq

All commands support --format json for piping to jq:

# Top 3 most expensive tests
agentv inspect show <result-file> --format json \
  | jq 'sort_by(-.cost_usd) | .[0:3] | .[] | {test_id, cost: .cost_usd, score}'

# Tests where token usage exceeds 10k
agentv inspect show <result-file> --format json \
  | jq '[.[] | select(.token_usage.input + .token_usage.output > 10000) | {test_id, tokens: (.token_usage.input + .token_usage.output)}]'

# Score distribution by suite
agentv inspect show <result-file> --format json \
  | jq 'group_by(.suite) | .[] | {suite: .[0].suite, count: length, avg_score: ([.[].score] | add / length)}'

# Tool usage frequency across all tests
agentv inspect show <result-file> --format json \
  | jq '[.[].trace.tool_calls // {} | to_entries[]] | group_by(.key) | .[] | {tool: .[0].key, total_calls: ([.[].value] | add)}'

# Find regressions > 0.1 between two runs
agentv compare baseline.jsonl candidate.jsonl --format json \
  | jq '.matched[] | select(.delta < -0.1) | {test_id: .testId, delta, from: .score1, to: .score2}'

Reasoning Patterns

When analyzing traces, think about:

1. Efficiency: Are tool calls/tokens proportional to task complexity? High tokens-per-tool may indicate verbose prompts or unnecessary context.

2. Error patterns: Do failures cluster by target, suite, or tool usage? Common patterns:

   • Tool errors → agent can't access required resources
   • High LLM calls with low tool calls → agent stuck in reasoning loop
   • Missing tool calls → wrong tool routing

3. Cost optimization: Identify tests with high cost but acceptable scores — can they use a cheaper model? Compare --group-by target stats.

4. Latency distribution: P50 vs P99 spread indicates consistency. Large spread means unpredictable performance — investigate P99 outliers.

5. Regression detection: After a prompt/config change, compare before/after. Mean delta > 0 is good, but check individual test regressions — a few large losses can hide behind many small wins.