Reader style: agent
The reader style of an agent handed a brief. It has three parts: who the agent is, five overrides with their grades, and what is not yet known.
Who
- Reader: an agent, a language model that is handed
START.md,memory/INDEX.mdand abrief.md, acts on a project with tools, and has its output checked by programs. - Expertise: broad and shallow, as in the language-model style. It knows the field and nothing of this project.
- Medium: a context window, with tools to read more files and to run commands.
- Reading mode: front to back once, then back to the brief to check what it was asked.
This style extends the language-model style. Everything that style and the standard say still holds. Only what this family's own evidence adds is listed.
Overrides
| Principle | Change | Why |
|---|---|---|
| overview first | strengthen: the brief, with its four headings, halves the failures of a cheap model against no brief; naming the checks adds nothing once the tests are hidden | 0 of 60 with no brief, 30 of 60 with the convention brief, 29 of 60 with the checks added, on hidden tests (Brief test 2026, round two; the brief: measured, descriptive, two tiers, three tasks; naming the checks: not supported) |
| one place per fact | strengthen: the start file carries one sentence giving the repository precedence over a stale instruction | a stale file was followed in 90.1% of choices; one sentence of precedence raised the share following the code from 4.2% to 91.4% (Xu 2026; measured, secondary) |
| size is a prompt | strengthen: the start file and the memory index each stay under a bound, and the rest is read on demand | the method's design; no study sets the length (convention, PROVISIONAL) |
| overview first | add: when several agents work on one project, a central check runs on their results | errors grew 17.2 times without central checking and 4.4 times with it (Kim 2025; measured) |
| (decisions) | add: the agent decides nothing that no program or person has checked | the blueprint's position and the method's decision 0003, from Vaccaro 2024 (convention) |
Sources: Brief test 2026 is the experiment of ai-augmented-agents. Round two is in its results and decision 0005. Round one was 5 of 10 with the checks listed, 1 of 10 rescored on hidden cases. The others are in the bibliography.
Thresholds
The bound is read from configuration, never typed into a file.
[agent_reader]
start_file_max_lines = 200 # PROVISIONAL: from the method's design, no study sets it
index_max_lines = 200 # PROVISIONAL: sameOpen questions
- Whether the brief effect holds beyond one task and one small model: not measured.
- Whether 200 lines is the right bound: not measured, and a vendor gives under 200 as practice only.
- Whether the precedence sentence helps when the stale file is short: not measured.
- Whether a central check pays for its cost with two agents rather than many: not measured here.