Arbor Agent Search

Arbor Search Agent

Use this for post-experiment related-work annotation. It is not for internal codebase search; use normal file tools for that.

Eligibility Gate

By default, only search nodes that are validated and effective:

• node status is done or merged;
• node has a numeric score;
• score improves over current trunk_score, or over baseline_score if trunk is unavailable.

Skip pending, running, unscored, and below-trunk nodes. This keeps novelty cost tied to merge candidates.

When To Search

• After RunExecutor returns a node with score > trunk_score.
• Before GitMergeBranch when novelty or related work affects the decision.
• After parallel executor results, annotate all sibling winners with SearchIdeaContextParallel.

Do not search trivial parameter tweaks, pure scale-ups, or internal codebase questions.

SearchAgent Task

The SearchAgent is a novelty scout. It does not implement, critique, or change the idea. It surveys prior work and writes a Markdown annotation to node.related_work.

Use 2-3 query angles:

• technique class;
• application domain;
• key mechanism.

For ML/NLP literature, use English academic queries with words such as paper, arxiv, or survey; include one original-language query when useful.

Hard Caps

• At most 2 search rounds.
• Visit up to 5 pages total.
• Keep the result short enough for a researcher to judge novelty quickly.

Final JSON Schema

When running an isolated SearchAgent, ask it to emit only:

{
  "summary": "2-4 sentences describing prior work.",
  "related_papers": [
    {
      "title": "Paper title",
      "url": "https://...",
      "one_line_relevance": "Why this is relevant."
    }
  ],
  "novelty_assessment": "novel | partial-overlap | prior-art-exists",
  "overlap_risks": "What specifically overlaps, or 'none'."
}

Render it into:

### Summary
...

### Related Papers
- [Title](url) - relevance

### Novelty
novel | partial-overlap | prior-art-exists - justification

### Overlap Risks
...

Background Behavior

Native Arbor usually runs SearchAgent in the background:

• SearchIdeaContext returns a dispatch message immediately.
• The coordinator continues IDEATE/DISPATCH work.
• The result lands on node.related_work later.
• SearchStatus reports in-flight searches.
• Shutdown waits for pending searches to flush.

When emulating in a host that lacks background tasks, run search synchronously and update related_work before merging.

Failure Handling

Failures are non-blocking. Write a marker such as:

[search-failed: <reason>]

Treat it as no information, not evidence of novelty.