---
title: "Discussion on \"Research Reports Your Agent Writes Need Sources Before Sentences\" | Hashnode"
description: "Discussion on \"Research Reports Your Agent Writes Need Sources Before Sentences\". Ask a coding agent for a research report and you usually get fluent prose first and evidence later — if at all. A single forum post becomes \"research shows.\" A number from 2023 is presented as the cur"
url: "https://hashnode.com/posts/research-reports-your-agent-writes-need-sources-before-sentences/6ac31d019cc029cefd045238"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Research Reports Your Agent Writes Need Sources Before Sentences](https://alapha888.hashnode.dev/research-reports-your-agent-writes-need-sources-before-sentences?utm_source=hashnode&utm_medium=feed)

Ask a coding agent for a research report and you usually get fluent prose first and evidence later — if at all. A single forum post becomes "research shows." A number from 2023 is presented as the cur

Oct 5 · 4 min read · alapha888.hashnode.dev  
Read the article

-   [#research](https://hashnode.com/tag/research)
-   [#ai](https://hashnode.com/tag/ai)
-   [#productivity](https://hashnode.com/tag/productivity)
-   [#writing](https://hashnode.com/tag/writing)

## Written by

[alapha888](https://hashnode.com/@alapha888)

@alapha888

  
[View profile](https://hashnode.com/@alapha888)

## Responses 4

Markdown supported · ⌘↵ to post

[Alrebro](https://hashnode.com/@alrebro)

[Oct 5](https://hashnode.com/posts/research-reports-your-agent-writes-need-sources-before-sentences/6ac31d019cc029cefd045238/comment/6ac351b5e96d19ea256d8037)

The "sources before sentences" framing nails it. The failure mode I keep seeing is agents treating search order as evidence order — whatever the search returns first becomes the report's structure, regardless of source quality. Your question-first step plus source tiering is a solid guardrail against that. One addition that pairs well with your freshness column: pin the "as of" date next to the question definition at collection time. Agents routinely present a 2023 number as current state simply because nobody told them when "now" is. A stale-data check right after tiering would close that loop.

[alapha888](https://hashnode.com/@alapha888)

[Oct 5](https://hashnode.com/posts/research-reports-your-agent-writes-need-sources-before-sentences/6ac31d019cc029cefd045238/comment/6ac351b5e96d19ea256d8037#reply-6ac356bb93fac4039279cf3a)

Thanks for this — "search order as evidence order" is a sharp way to name it. When retrieval rank quietly becomes the outline, whatever ranks first gets treated as foundational rather than just first-found. The fix I use is a source ledger keyed by claim, built before any outlining, so the structure comes from verified coverage instead of retrieval order.

[indiainfranotes](https://hashnode.com/@indiainfranotes)

[Oct 5](https://hashnode.com/posts/research-reports-your-agent-writes-need-sources-before-sentences/6ac31d019cc029cefd045238/comment/6ac34e38a255b24b3a4f536c)

The rule that two sources quoting the same origin do not count as independent is the one most agent research tools skip, and it is exactly how one press release turns into a consensus. Tracing each claim back to its origin URL before counting would make that check mechanical. Have you tried having the agent output a small claim-to-source table first and only write prose from that table?

iin1005h1228

[alapha888](https://hashnode.com/@alapha888)

[Oct 5](https://hashnode.com/posts/research-reports-your-agent-writes-need-sources-before-sentences/6ac31d019cc029cefd045238/comment/6ac34e38a255b24b3a4f536c#reply-6ac356f093fac4039279d482)

Appreciate this — one press release turning into a "consensus" through repeated quoting is exactly the failure mode the piece is about. Tracing each claim back to its origin URL before drafting helps, and one small addition: flag any claim as unverified in the draft itself when only a single origin exists, so it is obvious what still needs independent confirmation before publishing.
