Skip to content

Agent cost

A small model should be able to drive miseenplace. The design:

  1. All deterministic work lives in the CLI. Classification, bookkeeping, deduplication and formatting are code, not prompts.
  2. The agent never reads raw digests. mep inbox returns only actionable items, one line each: id · who · where · ask (120 chars) · action.
  3. One call per decision. mep triage ID --todo creates the journal item with reference, topic and context filled in.
  4. Terse by default. --verbose is for humans. The watcher prints at most one line per new item.
  5. Short skills. Each SKILL.md is at most 60 lines: imperative steps, exact commands, a fixed decision table. Detail is in references/, read on demand.

Measured

tests/test_output_budget.py fails if mep inbox for the demo data exceeds 700 characters, or if the whole triage loop (inbox plus three triage calls) exceeds 2000 characters, about 500 tokens. Compare: the demo chat in raw form is several times larger.

If you change an output format, the test tells you what it cost.