Mitchell Foutz All writing

Ten thousand coding-agent sessions: what the counts show

From 2026-04-15 to 2026-10-08 the Claude Code local stats cache recorded 10,202 sessions and 282,405 messages. Codex CLI wrote 2,195 session files from 2026-04-26 to 2026-10-26. Two days in August hold 59 percent of the Claude Code sessions and 5.5 percent of its messages. The counts record volume and timing; none of them measures the quality of the work.

Published
2026-10-28 (draft)
Topic
Agent engineering
Data
Claude Code stats cache, Codex session files
Period
2026-04-15 to 2026-10-26
Verification
Self-reported. The stats cache and the session files are local records that are not published, so the counts, medians and peaks on this page cannot be checked independently. data.json holds only the aggregates drawn in the charts.

What was counted and how

Two local records were read. Each tool writes its own record on the machine where it runs.

Nothing else was read. No transcript, prompt or session file content was opened. The figures below are sums of the daily entries or of the file counts, grouped by 14-day period, by ISO week (Monday start) or by calendar month.

The files define little. The stats cache does not state what it counts as a message or as a tool call, or whether a subagent run counts as a separate session. This article uses the counts as recorded.

Claude Code counts

Claude Code was active on 134 of the 177 calendar days from 2026-04-15 to 2026-10-08. On the median active day the cache recorded 7 sessions and 1,452 messages. Over the whole period there were 0.59 tool calls per message.

Two-week message totals peak at 41,284 in the period starting 2026-05-25. Grouped by ISO week instead, the peak is 30,170, in the week starting 2026-08-10. The peak depends on the grouping. The highest single day is 2026-09-09, with 17,728 messages in 14 sessions. The cache has no entry for any day in the week starting 2026-06-15.

Sessions and messages are distributed differently. August holds 7,345 of the 10,202 sessions (72.0 percent) and 26.1 percent of the messages. Most of the August sessions fall on two days: 2026-08-13 and 2026-08-14 record 6,001 sessions between them.

The per-model totals list 12 model identifiers and 206,473,081 output tokens. The largest share belongs to claude-opus-5, with 75,499,452 output tokens (36.6 percent). The file does not state the period those totals cover. Its daily per-model series starts later, on 2026-07-18.

Codex CLI session files

Codex CLI session files are dated on 131 distinct days, with a median of 7 files per active day. June holds 946 of the 2,195 files (43.1 percent). One day, 2026-06-25, holds 455 (20.7 percent). Inside the Claude Code window, 2026-04-15 to 2026-10-08, there are 2,133 Codex files.

A session file and a Claude Code session are different units, counted by different tools. This article does not add them together or compare their sizes.

What the counts cannot show

Quality of the output

Neither record says whether a session produced working code, passed its tests, was reviewed or was kept. A message count measures traffic between the user, the model and its tools. These counts support statements about volume and timing and no others.

Length of a session

Messages per session range from 10.0 in August to 1,171.8 in April, a factor of 116.6. The longest session in the cache has 4,361 messages and a recorded duration of 224.9 hours, starting 2026-05-13.

MonthSessionsMessagesMessages per session
April (from the 15th)1214,0621,171.8
May79970,80688.6
June68131,45346.2
July1,07831,15328.9
August7,34573,81110.0
September18954,547288.6
October (to the 8th)986,57367.1

The monthly ratio divides the messages recorded in a month by the sessions recorded in that month. The daily session counts sum to the file's session total. Inference: each session is counted on one day, and a session that runs across several days can add messages to a month in which it is not counted as a session. The ratio is an approximate indicator for that reason.

Automated sessions

The stats cache does not mark a session as interactive or scripted. The Codex file names do not either. On 2026-08-13 and 2026-08-14 the cache records 6,001 sessions and 15,439 messages, 2.57 messages per session. Inference: these are scripted runs that each send one prompt and exit. Without those two days the cache holds 4,201 sessions and 266,966 messages, 63.55 messages per session. Inference: the 455 Codex files dated 2026-06-25 are a scripted batch of the same kind.

Coverage

Controls used with these sessions

The controls used with these sessions are sandbox and review-gate scripts, published as darblinko/agent-controls. Neither record logs when a control ran, blocked an action or changed a result. This article makes no claim about their effect.