agentic-session-economicslisted

# Agentic Session Economics Keeping agent-session spend proportional to work delivered. Worked examples, recorded costs, and provenance: `examples.md`. ## The problem, named honestly Unmeasured file-loads and unscheduled doc updates silently dominate session budgets — not the feature code, not the reviewer passes. On the reference project, doc updates and gitignore checks consumed two sessions' worth of budget before anyone measured why (`examples.md` §1). **The general claim:** any artifact that grows monotonically — specs, practices docs, retrospectives, transcripts — eventually crosses a cost threshold where the workflow that built it stops being affordable. The response is some combination of (a) read less of it, (b) push the expensive work to a cheaper tier, or (c) batch the updates so the cost amortizes. Build the cost knobs in *before* they are forced on you. --- ## Measure before acting Before reading a file or starting a multi-step doc update, check the observable proxies: | Proxy | How to check | What it signals | |---|---|---| | File line count | `wc -l <file>` | >200 lines → consider tail-read; >400 → consider splitting | | Re-reads of one file in a session | Count loads of the same path | Context isn't holding; the model is substituting re-reads for working memory | | Compaction events | Harness notices or context-overflow warnings | Session is expensive to continue; a fresh start is likely cheaper | | Subagent token reports | Whatever your platform surf