Dont Stop Early: Scalable Enterprise Deep Research with Controlled Information Flow and Evidence-Aware Termination

🕒 Published (v1): 2026-04-27 20:31 UTC · Source: Arxiv · Venue: ACL · link

Why this paper was selected

Scalable enterprise deep research with evidence-aware termination; production architecture

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper presents an Enterprise Deep Research (EDR) architecture that addresses three systematic failure modes in multi-agent deep research systems: uneven coverage, context explosion, and premature stopping. The system decomposes queries into coverage-driven DAG-structured plans, localizes agent context via dependency-gated information sharing, and enforces explicit evidence-sufficiency termination criteria. It achieves state-of-the-art results on both an internal sales enablement benchmark and DeepResearch Bench.

Problem

Existing deep research systems fail in enterprise settings due to: (1) uneven coverage — agents address easily accessible topics while missing required analytical dimensions; (2) context explosion — sequential multi-agent pipelines accumulate irrelevant intermediate outputs, degrading downstream reasoning; and (3) premature stopping — agents terminate after surface-level evidence collection without verifying that sufficiency conditions are met. These failures produce inconsistent report depth and non-decision-ready outputs.

Method

The system operates in three stages:

1. Outline generation with reflection: A coverage-driven outline enumerates all required information objectives (e.g., for a win-card: customer background, org structure, pain points, competitive landscape) before any retrieval. A reflection step audits for missing fields or ambiguous sub-questions. Critically, no retrieval tools are called at this stage to avoid evidence-availability bias.

2. DAG-based planning with dependency-controlled context sharing: The outline is converted into a directed acyclic graph (DAG) of executable research steps. Each node specifies agent type (public researcher or internal agent), tool set, and dependency edges. Steps sharing no dependencies execute in parallel. Each agent receives only the outputs of its direct predecessors, not the full execution history, bounding context growth.

3. Evidence-based termination: At initialization, each agent defines explicit sufficiency criteria (e.g., "identify relevant teams, key roles, reporting relationships, and decision ownership"). The agent iterates a reasoning-and-action loop, gathering evidence and self-evaluating against these criteria until they are satisfied. A progress reflection step after each planning iteration compares collected outputs against the original outline and may trigger new steps or replanning.

Tool interfaces include web search (Tavily) for public information and MCP-based connectors for internal enterprise sources (CRM, Confluence/Highspot, Slack conversation search). Orchestration is implemented in LangGraph.

Key Contributions

• Coverage-first decomposition via outline generation with reflective auditing, performed without retrieval to avoid evidence-distribution bias.
• Dependency-structured plan DAG enabling parallel execution while preserving logical ordering and bounding per-agent context to direct predecessor outputs only.
• Step-level evidence sufficiency criteria that prevent premature termination by requiring agents to iteratively collect evidence until explicit conditions are met.
• Empirical demonstration that dependency-controlled context sharing reduces execution time from 222 to 47 minutes while improving report quality on a sales enablement benchmark.
• Ablation evidence isolating the contribution of each mechanism: agent termination criteria alone account for HAA drops from 82.09 to 73.67 when removed.

Results

Enterprise sales enablement (win-card generation, 10 accounts): - Our 4.1 (public search only): HAA 71.94, Coverage 3.99 — best among all DR baselines including Gemini DR and OpenAI DR - Our 4.1 (public + internal tools): HAA 82.09, Coverage 4.31, IIR 0.89 — leads all methods; IIR of 0.89 vs. 0.70 (DeerFlow) and 0.64 (ODR) - Sequential (no DAG) ablation: HAA 76.40, time 222 min vs. 47 min for full system - Removing agent termination: HAA drops to 73.67, public tool calls drop from 327 to 224 (premature stopping confirmed)

DeepResearch Bench: - Our 5.1: avg 53.40, comprehensiveness 54.85, insight 55.63 — best among all compared methods including Salesforce AIR (50.65), ThinkDepth.ai (52.43), Tavily Research (52.44) - Our 4.1: avg 47.02; Our 4.1m: avg 43.46

Limitations

• Win-card evaluation is specific to sales enablement; generalization to other enterprise task types requires new domain-specific evaluation protocols.
• Evaluation conducted on only 10 customer scenarios, limiting statistical power.
• Authentication, permission enforcement, and data residency compliance for enterprise tool access were not implemented in the research prototype.
• Template structure in the sales enablement task artificially reduces planning difficulty; performance on unstructured tasks relies more heavily on backbone model capacity (as seen in DeepResearch Bench results).
• Fixed number of planning/replanning iterations is a system hyperparameter; sensitivity analysis is not reported.

Relevance to Agentic AI / LLM Agents

This work directly addresses two canonical failure modes identified in multi-agent LLM systems research — premature termination and context explosion — with concrete architectural mechanisms rather than prompting heuristics. The evidence-based termination criterion pattern (define "done" before acting, iterate until satisfied) is a transferable design primitive applicable beyond deep research to any long-horizon agentic task. The DAG-structured execution with dependency-gated context sharing operationalizes a principled solution to the information bottleneck problem in multi-agent pipelines, complementing parallel lines of work on graph-based LLM reasoning and tool-use workflows. The ablation showing a 4.75× wall-clock speedup from parallelism (47 vs. 222 min) provides concrete evidence for why dependency-aware scheduling matters in production agentic systems.