Agent Teams Involved in the RESEARCH Stage of the ARS Pipeline

The RESEARCH stage is powered by two distinct agent teams: a core Deep-Research Skill Team comprising 13 specialized agents that execute the full research workflow, and a non-blocking Observer Team (the collaboration_depth_agent) that audits collaboration quality without halting execution.

The Imbad0202/academic-research-skills repository implements a sophisticated multi-agent system for academic research automation. Understanding the specific agent teams involved in the RESEARCH stage is essential for operators tracing how raw literature transforms into verified synthesis reports and INSIGHT collections.

Deep-Research Skill Team (13 Agents)

The primary workforce for Stage 1 RESEARCH is the Deep-Research Skill Team, defined in the architecture matrix at docs/ARCHITECTURE.md (lines 99-100). These agents collectively handle everything from research question formulation to final report compilation, operating under Pattern-Protection blocks (v3.6.7) to prevent prompt drift.

Research Design and Architecture

• research_question_agent – Drafts the initial RQ Brief, translating user intent into a structured research question with predefined scope and success criteria.
• research_architect_agent – Refines the methodology blueprint, selecting appropriate study designs and analysis frameworks based on the RQ Brief.

Literature Processing and Verification

• bibliography_agent – Consumes the user-supplied literature_corpus[] and applies the Four Iron Rules to ensure corpus-first integrity. This agent is the Phase 1 gatekeeper for all source material.
• source_verification_agent – Enforces provenance standards and validates citation chains against the source documents.

Synthesis and Analysis

• synthesis_agent – Generates the comprehensive synthesis report by integrating findings across the annotated bibliography.
• meta_analysis_agent – Performs statistical aggregation and heterogeneity testing when quantitative pooling is required by the methodology blueprint.

Quality Control and Ethics

• risk_of_bias_agent – Evaluates internal and external validity threats using standardized appraisal tools (e.g., ROB-2, GRADE).
• ethics_review_agent – Screens protocols and sources for ethical compliance, flagging sensitive data usage or consent issues.
• devils_advocate_agent – Stress-tests the emerging narrative by interrogating assumptions and identifying alternative interpretations.
• editor_in_chief_agent – Maintains stylistic consistency and structural coherence across all generated documents.

Compilation and Mentorship

• report_compiler_agent – Assembles the final RESEARCH deliverable, integrating all sub-components into the canonical output format.
• monitoring_agent – Records runtime metrics, token usage, and decision logs to the shared Material Passport for later audit.
• socratic_mentor_agent – Provides pedagogical guidance during the research process, ensuring the user understands methodological choices.

Observer Team: Non-Blocking Collaboration Audits

Separate from the execution team, the Observer Team consists of a single advisory agent:

• collaboration_depth_agent – Runs after decision-heavy checkpoints and integrity gates to log metrics on delegation intensity and cognitive vigilance. According to the architecture legend in docs/ARCHITECTURE.md (lines 199-205), this agent operates in advisory mode only and never blocks pipeline progress.

Workflow Integration and Execution Flow

The agent teams interact through a structured orchestration protocol defined in the skill-dependency graph:

1. User Checkpoint – The operator approves the RQ Brief + Methodology Blueprint (human gate 🧑 1).
2. Machine Pre-checks – API-tier-0 verification runs Levenshtein similarity checks, evidence-hierarchy grading, and anti-sycophancy filters on the brief and bibliography.
3. Parallel Agent Execution – Each Deep-Research agent receives the shared Material Passport and executes according to the dependency graph (lines 199-205).
4. Observer Logging – The collaboration_depth_agent records advisory metrics post-execution.
5. Stage Transition – Upon user confirmation, the pipeline advances to Stage 2 (WRITE).

CLI Interaction with RESEARCH Agent Teams

Operators interact with these teams through the ARS CLI. The commands trigger specific agents based on the current pipeline state:

# Initialize a full RESEARCH session (invokes research_question_agent and research_architect_agent)

ars plan --stage RESEARCH --mode full

# Inspect the annotated bibliography generated by bibliography_agent

ars lit-review --output bibliography.md

# Execute evidence synthesis (synthesis_agent + meta_analysis_agent)

ars synthesize --output synthesis_report.md

# Display the final compiled report from report_compiler_agent

ars show --stage RESEARCH --output research_report.md

These commands reference specification files in commands/ (e.g., ars-plan.md, ars-lit-review.md), which map CLI arguments to the underlying agent orchestration logic in deep-research/agents/.

Summary

• The RESEARCH stage employs 13 core agents in the Deep-Research Skill Team, ranging from research_question_agent to report_compiler_agent.
• An additional Observer Team (collaboration_depth_agent) provides non-blocking advisory feedback on collaboration quality.
• All agents operate under Pattern-Protection v3.6.7 and respect the Four Iron Rules when processing literature.
• Agent definitions are stored in deep-research/agents/*.md, with the architecture matrix documented at docs/ARCHITECTURE.md lines 99-100.
• Execution follows a user-heavy checkpoint model, with the Material Passport serving as shared state between agents.

Frequently Asked Questions

How many agents participate in the RESEARCH stage?

The RESEARCH stage utilizes 14 agents total: 13 in the Deep-Research Skill Team that perform substantive research operations, plus 1 observer agent (collaboration_depth_agent) that audits collaboration metrics. The 13 core agents are listed in the architecture matrix at docs/ARCHITECTURE.md lines 99-100.

What prevents the bibliography_agent from corrupting the source literature?

The bibliography_agent adheres to the Four Iron Rules defined in the corpus consumer protocol, ensuring corpus-first integrity. It processes the literature_corpus[] array without altering source documents, applying ground-truth isolation patterns specified in shared/ground_truth_isolation_pattern.md.

Does the collaboration_depth_agent block the pipeline if it detects issues?

No. The collaboration_depth_agent is explicitly designated as an observer that runs in advisory mode only. According to the architecture legend, it logs metrics regarding delegation intensity and cognitive vigilance but is architecturally prohibited from blocking progress through the RESEARCH stage.

Where are the individual agent prompts and specifications stored?

Each agent's prompt templates and behavioral specifications reside in deep-research/agents/ as Markdown files (e.g., research_question_agent.md, bibliography_agent.md). These files are protected by Pattern-Protection blocks to ensure reproducible agent behavior across pipeline executions.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}