Prompt clarity and quality
Assess how clearly the engineer states the goal and supplies useful context and constraints.
Agent Effectiveness · part of Developer Intelligence
See how your engineers work with agents. Make the best way the default.
Agent Effectiveness analyzes coding-agent sessions to help teams understand which practices produce useful, verifiable work. Compare session patterns, review coaching suggestions, and share practices across the team.
Claude CodeCodexCursorOpenCodeGitHub Copilot
Spotlights
Example sessions · Last 30 days
74 / 100
team effectiveness, example top quartile 81
12 of 64
engineers already scoring 80 or above
Plan before you build
19 engineers open feature work with a plan request. 34% fewer iterations, 6 example prompts.
+4 Feature development$1,900 / mo
Adoption 19 of 64 · discuss as a team practice
Name the files before the agent edits
Diffs 2.2× smaller when the prompt names the edit surface. 23% of feature sessions do.
+3 Feature development
Adoption 14 of 64
Run the suite before reporting done
56 of 97 bug-fixing sessions already do. The other 41 needed 2.9× more follow-up fixes.
Suggested CLAUDE.md practice+5 Bug fixing
Adoption 56 of 97 sessions · review the guide change
Illustrative practice suggestions. Confirm access and attribution settings with your organization's administrator.
Trusted by AI-forward enterprises
Agent mix
See the work behind your agent usage and cost.
Larridin brings coding-agent sessions, work patterns, token usage, and cost into one view. Session cost estimates use model pricing and usage telemetry; provider billing views show billed spend where a billing connection is available.
Work patterns · compare feature development, bug fixing, research, testing, and other work in the illustrated view
Session context · inspect agent configuration and linked pull requests alongside the session
Prompt cache accounted · cache creation and cache reads have different prices; inspect their contribution to estimated cost
Coverage matters · interpret comparisons alongside the sessions and coding tools available in the selected period
Agent mix
$21,480
estimated cost of analyzed sessions
612
analyzed sessions of 1,529 captured
$35
average per session, median $4
28%
of estimated cost in six sessions
Daily estimated cost by type of work
weekends lighter
$1,800
$1,200
$600
Aug 10
Aug 17
Aug 24
Aug 31
Feature developmentCodebase researchBug fixing & debuggingPrototypingTesting & qualityOther (ops, data, docs)
Type of workShareSessionsScoreSpend
Feature development11876$10,310
Bug fixing & debugging9766$3,437
Codebase research9679$2,148
Prototyping1971$1,933
Testing & quality16281$1,504
Ops & maintenance5870$859
Docs & knowledge3580$645
Data & analytics2773$644
Six sessions account for $6,014. The other 606 average $26. Review the six first.
Prompt cache
input tokens, all agents
91%
of input tokens read from cache in this example. Cache share varies across sessions.
$9,240
estimated difference from uncached input pricing in this example.
$412
estimated cache-creation cost across 62 example sessions.
Illustrative cost estimates use model list rates, including separate cache-creation and cache-read prices. Estimates can differ from invoices. Scores summarize the analyzed sessions in each group.
Sessions
Open any session. See what the agent did, what it cost, and how the pair worked.
Review available session traces, agent turns, token usage, estimated cost, and linked PRs. The score combines six dimensions: prompt clarity, session steering, sophistication, prompt quality, verification discipline, and task outcomes. Available dimensions carry equal weight.
Prompts, duration, cost and tokens · review time and usage signals alongside the session; the example shows elapsed and engaged time separately
A score with citations · inspect the practices and evidence behind a session score; the example below shows selected dimensions
Connected agents in one list · compare Claude Code, Codex, Cursor, and other connected agents with their linked PRs
How the session score works
Agent Traces
App: All
All · 1,529
Merged PR · 80
Needs attention · 166
ScoreSessionAppPRsEnded ↓
53Document router-prompt scoringClaude Code1Aug 31
77Larridin brand voice guidelinesClaude Code0Aug 30
57Optimize dbt CI pipelineClaude Code0Aug 29
81Git worktree setup automationClaude Code0Aug 29
67HubSpot lead generation syncClaude Code0Aug 29
70AI Impact dashboard filtersClaude Code0Aug 29
60GCP dbt job log analysisClaude Code0Aug 29
63Plugin and authentication flowClaude Code0Aug 28
57Local persistence for draftsClaude Code0Aug 28
67Spend Mapping UI migrationCodex0Aug 28
37Performance impact of the routerClaude Code0Aug 28
70AI Spend Intelligence exportCursor0Aug 27
Prompts
32
30h 23m open
Engaged
2h 40m
4 re-warms
Cost estimate
$37.54
Example model mix
Tokens
76.5M
99.5% cached
EvaluationRecommendationsDetails
Overall score
81
/ 100
▲ 9 vs team's August average of 72
A strong session. The engineer opened with an explicit goal and asked for a plan, held scope with five targeted corrections, and the agent verified its work before reporting completion.
Show full reasoning →
16 citations · 9 prompts, 7 agent turns
Selected dimensions · 3 of 6
team average
Prompt clarity77
Asks are scoped with an explicit session goal, constraints and an upfront plan request.
Prompt quality82
Each prompt combines the goal, a performance target and constraints for the CI speed-ups.
Session steering91
Catches drift quickly with targeted corrections and clear scope boundaries.
Scoring
Find useful practices and give your team guidance they can act on.
Session scores bring six equally weighted dimensions into a team view of agent use. Compare practices, inspect the sessions behind a pattern, and turn useful findings into guidance your team can review and adopt.
Adoption, outcomes and cost per practice · compare how a practice relates to outcomes in the sessions available to your team
Compare practices across your own sessions · look for patterns within comparable work and consistent reporting periods
Reviewable guidance · use a suggested practice to draft an agent guide change, review it with your team, and track later sessions
Effectiveness
74 / 100
▲ 2 vs August average of 72
12 of 64
engineers scoring 80 or above
6
dimensions in the effectiveness score
Example median · 68
Example top quartile · 81
your team · 74
Working well
Plan before you build
19 engineers · feature sessions finish in 34% fewer iterations
Name the files before the agent edits
23% of feature sessions · diffs 2.2× smaller
Targeted corrections instead of restarts
41 engineers · steering at top-quartile level
Tests run before reporting done
56 of 97 bug-fixing sessions · the fewest follow-up fixes in the org
Needs attention
Closing without a test run
41 of 97 bug-fixing sessions · 2.9× more follow-up fixes
Opening with "fix the bug"
33 sessions · 12 turns before a plan appears
Restarting instead of steering
27 sessions · a fresh session after each miss, context rebuilt every time
Idle gaps that rebuild the cache
62 sessions · $412 in cache rebuilds, no score effect
The score averages available ratings for prompt clarity, session steering, sophistication, prompt quality, verification discipline, and task outcomes, scaled to 100. Comparisons here are illustrative.
Example practice guidance
review with your team
1
Review
Run the test suite before reporting done
41 sessions closed without one, then needed 2.9× more follow-up fixes.
CLAUDE.md · apps/api+5 Bug fixing$640 / mo
2
Discuss
Ask for a plan before feature work starts
19 of 64 engineers already do. Share their pattern with the other 45.
Spotlight+4 Feature development$1,900 / mo
3
Explore
Compare models for codebase research
96 research sessions. A lower-cost model scored 78 against 79 at 42% less in this example.
Model comparison$900 / mo
Example guide change · apps/api/CLAUDE.md
## Before you report done+ Run `pnpm test --filter api`, paste the summary line.+ If a test fails, fix it or name it and say why.+ Never report done on a red suite.
Review suggested text
Discuss with team
For you
example coachingSix of your nine bug-fixing sessions this month closed without a test run. Teammates who run the suite first finish with 2.9× fewer follow-ups. Try /verify on the next one.
Example coaching based on session patterns. Access depends on your organization's roles and configuration.
Privacy
Make access clear from the start.
Agent sessions can contain sensitive context. Larridin scopes analytics by organization and role. Work with your administrator to confirm which session details are collected, who can access them, and the retention settings for your deployment.
Confirm session access · agree who needs session detail and who should use team summaries
Review collection settings · check collector and redaction settings before enabling transcript collection
Use team summaries · review team scores, work mix, and spend; confirm detailed access with your administrator
Access planning example
EngineerManagerAdmin
Transcript and promptsOwn sessionsPer policyPer policy
Session scoresOwn sessionsTeam rollupOrg rollup
Work mix and spendOwn sessionsTeam, by engineerOrg
Coaching suggestionsPersonalTeam guidanceOrg guidance
This table illustrates an access policy to review with your administrator. It does not describe default permissions; available controls depend on your deployment.
A closer look
How the Agent Effectiveness Score works
Each scored session is evaluated across six skills. The composite is the mean of the available 1–5 dimension scores, converted to a 0–100 scale. Unobserved dimensions are excluded rather than treated as zero.
Steering and sophistication
Assess how the engineer guides the session and uses the agent’s capabilities to work through the task.
Verification and outcomes
Assess whether the work is checked and what the session accomplished. Follow the session evidence behind the score.
| Coding agents | What to check before comparing |
|---|---|
| Claude Code, Codex, Cursor, OpenCode | Confirm session capture and attribution for the versions your organization uses. |
| GitHub Copilot, Devin, Gemini CLI, Antigravity | A tool appearing in the catalog does not imply identical session coverage. Review captured and unscored sessions. |
Questions about Agent Effectiveness
What does the Agent Effectiveness Score measure?
It measures observed skills in a coding-agent session: prompt clarity, session steering, sophistication, prompt quality, verification discipline, and task outcomes. It is not an overall rating of an engineer’s productivity.
How should we compare scores?
Compare sessions with similar work types and coverage using the same rubric. Read the evidence and number of scored sessions alongside any average. Differences can reflect task difficulty, capture coverage, or work mix.
Who can see transcripts and session data?
Session data is organization-scoped and access depends on role, enabled tools, and sharing configuration. Review access controls with your administrator before capturing sensitive sessions. Model-provider retention and Larridin session storage are separate controls.
How should I interpret the product examples?
The product views on this page use illustrative data to explain the metrics and workflows. For metric definitions, sample calculations, and assumptions, see our measurement methodology.
Start this week.
Connect your coding agents and review the sessions available to your team. Try it on your own repos, or get pricing for your org.