Quality & Reference Materials

Adding a dedicated critic agent that reviews and scores output before presenting it increases accuracy from ~45% to ~55% across three rounds.

Research from the PaperBanana framework (Google) showed that adding 1-3 rounds of critique from a dedicated agent improved output quality by roughly 10%. The critic evaluates on specific dimensions — faithfulness, conciseness, readability, aesthetics — and either approves or sends the work back for revision. This feedback loop is the single biggest quality lever in agent team design. Without it, you get first-draft quality. With it, you get iterated, polished output.

Giving agents reference examples teaches structure better than any prompt instruction. Monkey see, monkey do.

Instead of writing a 500-word prompt describing the exact output format you want, show the agent an example of a good output. Drop reference files in a dedicated folder and point the research agent to them. The agent extracts patterns — style, structure, tone, format — more reliably from examples than from abstract descriptions. This is the retrieval step in the PaperBanana pipeline, and it's foundational.

Every agent should know its exit criteria — what output format, what quality bar, what signals "task complete."

Without explicit completion criteria, agents either under-deliver (stop too early) or over-deliver (keep iterating forever). For a research agent: "Output a JSON style brief with keys: palette, composition, typography, mood, references." For a critic: "Score each dimension 1-10. If all scores are 7+, approve. Otherwise, return specific improvement instructions." Clear exit criteria make agents predictable and their outputs consistent.