Tokens burned per week
Monthly token burn per dev
GB of logs stored
Token burn per PR
Median cost per line changed
% of sessions showing context rot
Claude Opus
Codex
Success rate: one-shot vs planning
Baseline
One-shot
Median initial context tokens
Claude Opus
Codex
Skills + MCPs as % of initial context
Loaded skill + MCP tokens that go unused
Est. weekly tokens wasted on unused skills + MCP (log scale)
- 1Plan with the agent
- 2Review the plan in detail
- 3Agent implements
- 4Review the tests it built
- 5Push to CI → automated adversarial review
- 6Guided human review
Cadence
Better AI coding for cheaper token bills.
teamcadence.ai
Coding agent logs -> save money.
First month free
CDNC-AI-ENG-MEL
Meta: Tokenmaxxxxxing
More adoption! Use the coding agent for everything.
ParallelismSubagents/goalDynamic workflowsRalph loops
001
Meta: Tokenefficiency
Everyone's being moved off subscriptions.
- Instantly 6x - 10x spend.
- Orgs are budgeting $100k/dev/year for tokens.
- Similar to cloud - devs control huge spend.
002
Real coding logs
We have terabytes of coding logs from our customers.
We do aggregated, anonymised analysis over them.
- Vibe check.
- Actionable insight about better AI coding.
003
Outcomes matter now
One imperfect metric is token burn per PR.
- Bubble size is avg PR size.
- Smart teams have relatively big PRs at relatively low cost.
- High volume generally improves outcomes.
004
Smaller changes better?
Every session carries a big fixed cost: loading context, exploring, planning.
- Smaller changes way more expensive per line.
- But obviously easier to review and work with.
- Action? find the sweet spot.
005
Bigger changes worse?
Measured using dev frustration rate and tool call failure rate.
- Claude is worse as sessions get bigger.
- Past 100m tokens it always gets worse.
- But fine in "normal" usage ranges.
- Action? find the sweet spot.
006
Plan vs one-shot
Everyone plans; I guess here's why.
- Outcomes are much worse when one-shotting.
- Action? planning is the job, get good at it.
007
Initial context is rising
System prompt, skills, MCPs.
- Claude just keeps increasing.
- Claude system prompt is much bigger.
008
Skills and MCPs
Skills are big.
- Skills are install and forget.
009
Unused Skills or MCPs
Most aren't used.
010
Pure token wastage
Huge across our whole client base.
- Action: time for a spring clean.
011
Simple recommended workflow
Balance up human and agent.
- Review the plan.
- Review the tests.
- Guided review of the code and decisions.
012
Cadence
Visibility for eng leaders and coaching for devs.
- Token and cost tracking across tools.
- Per-developer tailored recommendations about levelling up.
- PR reviews bringing together logs and code.
Hit me (Dave Slutzkin) up on LinkedIn for a chat, or use the coupon code.