How Scoring Works

Every session score is computed locally by the CLI from your session's telemetry, and the server independently recalculates it from the submitted dimension scores on every submission. If the recalculated score differs from the submitted one by more than 1 point, the submission is rejected. There is no hidden adjustment and no server-side tuning — this page is the whole formula.

The five dimensions

Your session score is a weighted sum of five dimension scores, each 0–100:

score = savingsRate * 0.35
      + cacheEfficiency * 0.25
      + wasteControl * 0.20
      + summaryCoverage * 0.15
      + volumeBonus * 0.05

The result is clamped to 0–100 and rounded.

1. Savings rate — weight 0.35

The fraction of raw output tokens that AgentOps saved by returning summaries instead of full output:

value = totalSavedTokens / totalRawTokens
score = min(100, (value / 0.6) * 100)

A 60% savings rate earns the full 100. This is the largest weight because token savings is the core job of the tool.

2. Cache efficiency — weight 0.25

The local cache hit rate:

value = totalCacheHits / totalCacheLookups
score = value * 100

If there were zero cache lookups, this scores 0.

3. Waste control — weight 0.20

Penalizes detected waste patterns (for example, re-running the same expensive command repeatedly) relative to total commands:

wasteRatio = wastePatterns / totalCommands
value = max(0, 1 - wasteRatio * 5)
score = value * 100

One waste pattern per 5 commands zeroes this dimension out.

4. Summary coverage — weight 0.15

The fraction of command families (test, build, typescript, eslint, etc.) where more than half the commands got a compact summary:

value = familiesWithSummaryRate>0.5 / totalFamilies
score = value * 100

5. Volume bonus — weight 0.05

A small logarithmic bonus for real usage, capped at 100:

value = min(1, log2(totalCommands + 1) * 15 / 100)
score = value * 100

About 100 commands in a session maxes this out. At 5% weight it can never carry a score — it exists so a 2-command session doesn't outrank a 200-command one on ratios alone.

Grades

Grades map directly from the final score:

• A+ — 95 and above
• A — 90–94
• A- — 85–89
• B+ — 80–84
• B — 75–79
• B- — 70–74
• C+ — 65–69
• C — 60–64
• D — 50–59
• F — below 50

A session with zero commands or zero raw tokens scores 0 (F) and is flagged as limited data. Sessions with fewer than 3 commands are also flagged as limited data.

Anti-gaming validation

The server does not trust the submitted score. On every submission it checks:

• the score recalculated from the five dimension scores matches the submitted score within 1 point
• savingsRate.value matches totalSavedTokens / totalRawTokens within 0.01
• cacheEfficiency.value matches totalCacheHits / totalCacheLookups within 0.01
• totalSavedTokens does not exceed totalRawTokens
• the per-family token breakdown sums to within 10% of totalRawTokens

Any mismatch rejects the submission with a specific error. The validation source is public: server/src/validation.ts.

Seeing your own breakdown

agentops score

prints the per-dimension values, scores, and weights for your latest session. agentops score --session <id> scores a specific session. The scoring source is public: src/core/scoring/sessionScorer.ts.