notes from the field

what a whole bunch of CC session logs reveal about how to get more out of the agent

Dave Slutzkin · CEO, Cadence · May 2026

intro

this is what we're seeing right now as of May 2026 - we have a bunch of session logs and we do a lot of analysis over them.

we also use a lot of coding agents ourselves, mostly codex right now sorry gang but with a lot of claude too.

html is cool

Using Claude Code: The Unreasonable Effectiveness of HTML — Thariq (@trq212)

don't use plan mode

PLAN MODE CONSIDERED HARMFUL

watch out for long sessions

Frustration rate
Claude Codex
1.2% 0.9% 0.6% 0.3% 0% increasing session length →
Tool failure rate
Claude Codex
8% 6% 4% 2% 0% increasing session length →

Know your model! Claude gets worse as sessions get longer. Interaction patterns with Claude are quite different to interaction patterns with Codex.

errors

been a bad month for CC, from our data

Error rate
8% 6% 4% 2% 0% last 30 days

rollback

Rollback seems to be getting increasingly powerful. people are interrupting models more.

But our preliminary data suggests this breaks your cache, so it's a juggle between context size/correctness, and cost/tokens.

cache busting

Cache expiry analysis — cache expiry is the single biggest waste category

Worth reading the original reddit post, but the point here is that your cache matters and idle time of sessions matters.

If you care about tokens/cost then pay attention.

creativity

Placeholder: the agent surprises you when you give it room to. Lean into the weird outputs — that's where the new ideas come from.

Open Collider

cadence

Cadence org overview
Cadence my work

discount code 2 months free: CC-MELB-MAY