AutoSearch — deep iterative research

Today’s standard search is single-shot: you provide parameters, you get back top results, that’s the search. AutoSearch is the multi-round upgrade — a research orchestrator that runs multiple searches with progressive refinement, converges when results stabilize, and returns a deeper, more grounded answer.

What this will let you do

• Describe a goal at a higher level than today’s standard search lets you. Examples:
  • “Find me three uncrowded sub-niches in pet care that I could enter at $30+ margin.”
  • “Identify the trend within home-fitness that’s most likely to grow over the next 6 months.”
  • “Compare my current top SKU to its closest 5 competitors and tell me what they’re doing differently.”
• Let the system iterate. AutoSearch runs an initial broad search, scores the results, identifies what’s missing or weak, and runs follow-up searches automatically.
• Watch convergence. A live progress view shows you each iteration as it completes, with the system explaining what it learned and what it’s trying next.
• Stop early when you want. Convergence is automatic, but you can intervene at any iteration.

How it’ll work (preview)

AutoSearch is a state-machine orchestrator on top of the same search engine. The flow:

1. Goal extraction. Claude reads your high-level goal and extracts a research plan with sub-questions.
2. Initial broad search. First iteration runs a wide search to map the territory.
3. Result quality score (RQS). The system scores how well the current results answer the goal. Low RQS → keep iterating.
4. Plan refinement. Claude proposes the next search to fill gaps.
5. Iterate until convergence. Repeat until RQS plateaus or hits a max-iteration cap.
6. Final synthesis. Claude summarizes the multi-round findings into a single answer with all underlying searches linked.

You’ll see a streaming progress view (Server-Sent Events) showing each iteration as it happens. The chat surface stays open so you can ask follow-up questions about what AutoSearch found.

Why iterative is harder than it sounds

Single-shot search returns whatever the first cut produces. AutoSearch has to:

• Know when to stop. (Convergence detection — too early misses opportunities; too late wastes credits.)
• Avoid redundant searches. (Don’t search the same niche twice with slightly-different params.)
• Handle the case where the goal is unanswerable. (Some goals don’t have a good answer; AutoSearch should say so rather than spinning.)
• Stay within credit budget. (We’re not letting AutoSearch silently burn 50 credits.)

The repo has these patterns spec’d out in detail in RUN3-PRD.md — convergence is a research-grade problem and we’re treating it as one.

Estimated cost

AutoSearch will be credit-budgeted. You set a maximum (e.g., 10 credits); the orchestrator stops at the budget regardless of convergence. Typical budgets:

• Quick AutoSearch: 3-5 iterations, ~5-7 credits.
• Standard AutoSearch: 5-8 iterations, ~10-15 credits.
• Deep AutoSearch: 8-15 iterations, ~20+ credits.

You’ll see the budget remaining in the progress view; you can extend or stop early.

Status

• Backend orchestrator: designed in RUN3-PRD.md, in active development.
• State machine + convergence detection: in development.
• SSE streaming for progress: planned (the same stream pattern as today’s chat).
• Frontend UI: part of Run 4 (frontend in hilal-software).
• Beta: planned for the Run 3 release.

Related articles