Post 0 · figure options · work Mac, Dec 2025 – Oct 2026

Where my agents' context comes from

Every tool call in 1,777 Claude Code sessions from the work laptop, bucketed by where the information came from. "Volume" is the text the tool returned (≈ tokens, chars ÷ 4). "Reach" is the share of sessions that touched the source at all. Only paths and tool names were read, never contents.

Option A

Donut: share of volume

Says"Here's the split." Familiar, readable in two seconds.
MissesEverything about how it gets there. Small slices (meetings) vanish, though they matter.
Option B

Waffle: 200 squares of context

SaysSame as the donut, but easier to compare and friendlier on a blog. Each square is 0.5%.
MissesSame blind spot: shows proportion, not flow or frequency.
Option C

Volume next to reach, by kind of session

SaysThe brain sessions read Slack and meetings; workers read the brain and code. Toggle to see the two jobs.
MissesNeeds the reader to toggle and compare. More analytic than inspiring.
Option D

Flow: which session pulls each source

SaysThe brain sessions do the collecting (Slack, meetings, docs), write it into the vault, and workers read the vault. Supports "the context was already collected".
CaveatThe vault → workers link is the brain-reads in worker sessions; "collected into the vault" is inferred from where the fetches happen, not measured.
Option E

Reach vs volume

SaysSeparates "touched in most sessions" from "brings a lot of text". Meetings sit low on volume but high on reach in brain sessions.
MissesLog scale and two axes: the most work for a reader skimming a blog.
Option F

What agents read, and how much got done

SaysTop: the mix of what agents read each week (average of that week and the two before); the brain arrives and immediately takes a big share. Bottom: MRs I opened (line: same 3-week average; faint bars: actual weeks), roughly 6 a week before the brain and 16 after. Three unrepresentative weeks are removed from both panels.
CarefulThe MR climb starts about three weeks before the brain (workers were ramping up too), so this is "together with", not "because of". Counted from MRs opened through Claude Code sessions, which by then was how I opened all of them; the chart starts at the first one.
Method. Read/Grep/Glob paths under the vault → brain (by folder); other paths → code. Shell commands bucketed by tool (git/glab → code, kubectl/aws/docker → prod, brain-search → brain). MCP connectors by name. Agent's own scratch files, file edits and task plumbing excluded. Subagent transcripts count toward their parent session. Groups. Meetings = Krisp + Glyphic connectors (meeting notes already in the vault count as brain). Prod & data = Datadog, kubectl/aws, internal dashboards, ClickHouse. Own history & skills = past session transcripts, skills and agent config.