695 Github Patterns Df38fb37

GitHub Research: Existing Research Agent Patterns

Research Date: December 9, 2025 Methodology: Systematic search of GitHub repositories for Claude Code research agents, awesome lists, and multi-agent research workflows.

Awesome Lists Found

PATTERN 1: Comprehensive Research Agent Team Architecture

Source: yuz207/claude-agents-research-team Repository: yuz207/claude-agents-research-team (⭐ Not starred yet - active development) Last Updated: 2025-12-07 Type: Production Claude Code agents for multi-phase research workflows

Team Structure (Verbatim)

research_team:
  core_scientists:
    - ai-research-lead      # PI: Designs, experiments, analyzes
    - ml-analyst           # Validates, profiles, optimizes
    - experiment-tracker   # Documents everything
  research_engineers:     # Called when needed for implementation
    - architect           # Designs complex system architectures
    - developer          # Implements new architectures cleanly
    - debugger           # Diagnoses training failures
    - quality-reviewer   # Pre-production validation
  potential_future_members:
    - technical-writer     # Manuscript preparation, literature review
    - science-critic      # Devil's advocate, challenges assumptions

Communication Architecture (Verbatim)

Human (You)
    ↓
Claude Code (Executive Interface)
    ↓
ai-research-lead (Principal Investigator)
    ↓ [Delegates to]
    ├── ml-analyst (Senior Empirical Analyst)
    ├── experiment-tracker (Research Secretary)
    └── [When needed] Research Engineers
        ├── architect (Complex designs)
        ├── developer (Clean implementations)
        └── debugger (Failure diagnosis)

Key Features:

• Stateless agents: Each invocation starts fresh, no memory of previous calls
• File-based context sharing: Uses agent_notes/[timestamp]_[workflow_name]/ directories
• Request-based coordination: Agents request Claude Code to invoke other agents
• Clear role separation: PI coordinates, analysts validate, engineers implement
• Structured outputs: All agents return typed findings/results

Agent Role Definitions (Extracted)

ai-research-lead:

• Principal Investigator & Lead Data Scientist
• Generates and tests hypotheses
• Performs statistical analyses and modeling
• Coordinates specialist agents
• Makes research recommendations

ml-analyst:

• Empirical model evaluation
• Statistical validation with evidence-based rigor
• A/B test analysis
• Root cause analysis
• Production monitoring
• Challenges PI's hypotheses with data

experiment-tracker:

• Documents all experiments
• Records decisions and context
• Preserves research artifacts
• Maintains searchable records
• Takes meeting minutes

architect:

• Designs complex implementations for novel architectures
• System design specifications
• Never implements, only designs

developer:

• Implements research ideas with production-quality code
• Writes comprehensive tests
• Implements architectures designed by architect

debugger:

• Root cause analysis of training/implementation failures
• Investigates NaN losses, gradient explosions
• Evidence-based diagnosis only

Critical Communication Rules (Verbatim)

1. Agents are stateless - Each invocation starts fresh, no memory of previous calls
2. Agents can't see chat history - They only see what's in the Task tool prompt
3. Agents can't directly invoke each other - Must request Claude Code to do it

SOLUTION: Request-based coordination with COMPLETE context:

1. Agents request Claude Code to invoke other agents
2. Must provide ALL necessary context (data, findings, code, etc.)
3. Surface ALL findings visibly in output for human and Claude Code
4. Claude Code orchestrates invocations and maintains context

PATTERN 2: Chief of Staff Workflow Orchestration

Source: yuz207/claude-agents-research-team - chief-of-staff.md File Path: /chief-of-staff.md

Mandatory Context Preservation Protocol (Verbatim)

BEFORE Invoking ANY Agent, MUST execute these steps:

1. Create Workflow Directory

   agent_notes/[timestamp]_[workflow_name]/
   Example: agent_notes/20250117_research_validation/
   

2. Write Agent Context File (MUST include ALL of these):

   # Agent Context: [Task Name]
   Created: [ISO timestamp]
   Workflow Phase: [1/2/3]
   
   ## Overall Workflow Plan
   [The complete multi-phase plan I'm executing]
   
   ## Delegation Reasoning
   [WHY this specific agent for this task - based on their expertise]
   
   ## Current Phase Objective
   [What this specific agent needs to accomplish]
   
   ## Previous Agent Findings (if applicable)
   - [Key discoveries from earlier phases]
   - [Critical patterns identified]
   - [Important decisions made]
   
   ## Current State
   - [What's been completed]
   - [What's in progress]
   - [Known issues or blockers]
   
   ## Critical Requirements
   - [Specific output format needed]
   - [Validation requirements]
   - [Success criteria]
   
   ## Hypothesis & Research Directions
   - [Current hypotheses being tested]
   - [Promising patterns to investigate]
   - [Alternative approaches if primary fails]
   
   ## File Outputs
   - Your output file: agent_notes/[timestamp]/phase[N]_[agent]_[detailed_purpose].md
   - Previous phase files: [list for reference]
   - IMPORTANT: APPEND to your file, don't overwrite
   
   ## For Agent to Document:
   - [ ] Initial analysis plan
   - [ ] Key findings and discoveries
   - [ ] Decisions made and rationale
   - [ ] Recommendations for next phase
   - [ ] Final summary of work completed
   

Parallel vs Sequential Execution Rules (Verbatim)

Prefer parallel when possible for speed:

• Independent analysis needed (discovery phase)
• Multiple aspects to check (security + performance + design)
• Time is critical
• No dependencies between agents

Use sequential when:

• Output dependencies exist (developer → debugger to verify)
• Building on findings (research-lead → ml-researcher to validate)
• Iterative refinement needed (developer ↔ debugger cycles)
• Each phase informs the next

Agent Invocation Pattern (Verbatim)

Task(subagent_type="research-lead", prompt="[context + task]")

When to include "ultrathink":

• Complex debugging or root cause analysis
• Statistical analysis or hypothesis testing
• Architecture design decisions
• Performance optimization
• Any task requiring deep reasoning

PATTERN 3: Academic Research Workflow - 12-Agent System

Source: adrianstier/research-agent File: ai_research_workflow_agent_template.md Repository: adrianstier/research-agent Last Updated: 2025-11-20 Type: Academic research pipeline from hypothesis → manuscript → submission

Complete Agent Suite (Verbatim)

1. Orchestrator Agent

• Role: Coordinate all specialized agents to move research project from idea to submission-ready manuscript
• Process:
  1. Read the Research Project Brief
  2. Determine which stage the project is at (framing, EDA, modeling, writing)
  3. Select which agent(s) to activate next
  4. Draft precise task instructions for each agent
  5. Review outputs for consistency and traceability
• Outputs: Project roadmap, status reports, task instructions to other agents

2. Research PRD Agent (Problem Definition)

• Role: Convert high-level research question into precise, structured PRD
• Inputs: Research Project Brief, data description, project notes
• Outputs:
  • Umbrella question + nested questions
  • Hypothesis tree with expected directions
  • Causal diagram (described in text)
  • Variable definitions (predictors, responses, random effects)
  • Study design summary
  • Planned endpoints and metrics
  • Planned primary vs secondary analyses
  • Risk & confounder list

3. Literature & Conceptual Framework Agent

• Role: Synthesize domain knowledge and theoretical frameworks
• Outputs: Literature synthesis, conceptual models, theoretical basis for hypotheses

4. Data QA & Cleaning Agent

• Role: Quality assurance on dataset
• Outputs: Data quality report, cleaning decisions, data dictionary

5. EDA (Exploratory Data Analysis) Agent

• Role: Systematic exploration of data structure
• Outputs:
  • EDA plan organized by variable
  • Catalog of 15-25 recommended plots with code templates
  • Interpretive notes tied to hypotheses
  • Flagged issues (surprises, violations)

6. Modeling Agent

• Role: Design and document statistical analysis plan
• Outputs:
  • Modeling PRD with model specifications
  • Primary & secondary model specifications
  • Sensitivity analysis grid
  • Code skeletons for R/Python
  • Diagnostics and fit checks

7. Figure Factory Agent

• Role: Design all main and supplementary figures
• Outputs:
  • Figure PRD (F1–F?, S1–S?)
  • Panel designs and aesthetics
  • Caption drafts
  • Code templates for each figure

8. Scientific Writer Agent

• Role: Transform analyses into polished manuscript sections
• Outputs:
  • Draft Introduction (literature + hypotheses)
  • Draft Methods (reproducible detail)
  • Draft Results (effect sizes, not p-values)
  • Draft Discussion (interpretation + literature)

9. Reference Agent

• Role: Manage citations and references
• Outputs:
  • Clean reference list
  • Citation consistency check
  • Formatted bibliography

10. Reviewer Agent

• Role: Internal review from multiple personas
• Outputs:
  • Reviewer reports (Reviewer 1, 2, Statistical)
  • Consolidated revision suggestions

11. Submission Agent

• Role: Prepare all submission materials
• Outputs:
  • Cover letter tailored to journal
  • Highlights/key points
  • Author contribution statement
  • Data & code availability statements
  • Compliance checklist

Research Project Brief Template (Verbatim)

# Research Project Brief

## 1. Project Title
[Working title]

## 2. Umbrella Question
[High-level research question]

## 3. Nested Questions & Hypotheses
- Q1:
  - H1a:
  - H1b:
- Q2:
  - H2a:
  - H2b:

## 4. System & Concepts
- Focal system:
- Key taxa / entities:
- Scales (space, time):
- Core mechanisms of interest:

## 5. Data Overview
- Data source(s):
- Observational vs experimental:
- Response variables:
- Predictor variables:
- Random effects / grouping:
- Known limitations / quirks:

## 6. Target Outlet & Constraints
- Target journals:
- Word limits:
- Figure/table limits:
- Any style constraints:

## 7. Deliverables
- Main manuscript
- Supplementary information
- Core figures (F1–F4 or more)
- Supplementary figures/tables
- Reproducible code repo
- Data & code availability statements

## 8. Known Risks & Confounders
[List potential issues: bias, missing data, design limitations, etc.]

PATTERN 4: Multi-Agent Research Pipeline with Pydantic AI

Source: aldiakhou/codex-main - x.py File: x.py (Complete implementation) Repository: aldiakhou/codex-main Type: Python-based multi-agent research pipeline using Pydantic AI + OpenAI

Architecture Pattern (Verbatim)

User ─▶ System ─▶ LeadResearcher ─▶ Subagents (A,B,...) ─▶ Memory ─▶ CitationAgent ─▶ System
           ▲                            │                                        │
           └─────────────── iterative research loop ◀─────────────────────────────┘

Core Data Structures (Verbatim)

class Synthesis(BaseModel):
    """LeadResearcher synthesis output from findings."""
    executive_summary: str
    findings: list[Finding]
    gaps_or_open_questions: list[str] = Field(default_factory=list)
    recommend_next_iteration: bool


class Finding(BaseModel):
    subtask_id: str
    aspect: str
    summary: str
    # link to citations gathered while researching that subtask
    citations: list[str] = Field(default_factory=list)


class Plan(BaseModel):
    """Plan created by LeadResearcher for this iteration."""
    steps: list[Subtask] = Field(default_factory=list, description="Sub-tasks to execute now.")
    continue_research: bool = Field(
        description="True to continue iterative loop after current steps are done."
    )
    rationale: str


class Subtask(BaseModel):
    id: str = Field(description="Unique ID for the sub-task, short (e.g. 'A' or 'B').")
    aspect: str = Field(description="What this sub-agent should research.")
    target_depth: str = Field(
        description="Depth like 'quick scan', 'deep dive', 'verify claims'."
    )


class CitationRequest(BaseModel):
    """What we give the CitationAgent."""
    draft_report_markdown: str
    bibliography: dict[str, str]  # (url -> "Author, Title, Year" etc.)


class FinalReport(BaseModel):
    """What the CitationAgent returns after inserting citations."""
    markdown_with_citations: str

Iterative Research Loop (Verbatim)

async def research_pipeline(user_query: str) -> str:
    """
    End-to-end:
      1) Create dependencies (HTTP client, Memory)
      2) Create agents (Subagent, LeadResearcher, Planner, CitationAgent)
      3) Iterate: plan -> delegate -> synthesize, until plan says stop (or cap)
      4) Ask CitationAgent to insert citations and return the final report
    """
    # ... initialization ...

    all_findings: list[Finding] = []
    max_loops = 3  # safety cap
    iteration = 0

    while iteration < max_loops:
        iteration += 1

        # --- Plan the next iteration
        plan_res = await planner.run(
            f"User query:\n{user_query}\n\nMemory:\n{json.dumps(deps.memory.all(), indent=2)}"
        )
        plan = plan_res.output
        # Persist the plan to memory for traceability
        deps.memory.save(f"plan_iter_{iteration}", plan.model_dump_json())

        # --- Execute plan by delegating to sub-agents
        iter_findings: list[Finding] = []
        for step in plan.steps:
            try:
                finding = await lead.tools["delegate_to_subagent"](
                    subtask_id=step.id, aspect=step.aspect, depth=step.target_depth
                )
                iter_findings.append(finding)
            except UnexpectedModelBehavior as e:
                iter_findings.append(
                    Finding(
                        subtask_id=step.id,
                        aspect=step.aspect,
                        summary=f"Failed to research due to error: {e}",
                        citations=[],
                    )
                )

        all_findings.extend(iter_findings)

        # --- Synthesize for this iteration
        synth_prompt = (
            "Synthesize findings for this iteration.\n"
            f"User query: {user_query}\n"
            f"Iteration: {iteration}\n"
            f"Findings JSON:\n{json.dumps([f.model_dump() for f in iter_findings], indent=2)}\n"
        )
        synth = (await lead.run(synth_prompt, deps=deps)).output

        # Save synthesis snapshot
        deps.memory.save(f"synthesis_iter_{iteration}", synth.model_dump_json())

        # Exit/continue?
        if not plan.continue_research or not synth.recommend_next_iteration:
            break

    # --- Final drafting
    draft_md_lines: list[str] = [
        f"# Research report",
        "",
        f"**User query**: {user_query}",
        "",
        "## Executive summary",
    ]

    # Aggregate the last synthesis
    last_synth_json = deps.memory.recall(f"synthesis_iter_{iteration}") or ""
    if last_synth_json:
        try:
            last_synth = Synthesis.model_validate_json(last_synth_json)
            draft_md_lines.append(last_synth.executive_summary)
        except ValidationError:
            pass

    draft_md_lines += ["", "## Findings"]
    for f in all_findings:
        cites = " ".join(f.citations) if f.citations else ""
        draft_md_lines.append(f"- **[{f.subtask_id}] {f.aspect}** — {f.summary} {cites}")

    draft_md = "\n".join(draft_md_lines)

    # --- Citation insertion
    biblio: dict[str, str] = {}
    for k, v in deps.memory.all().items():
        if k.startswith("[cite:"):
            biblio[k] = v

    cite_req = CitationRequest(draft_report_markdown=draft_md, bibliography=biblio)
    final = (
        await citation.run(
            "Insert citations into the draft and add a 'References' section at the end.",
            deps=deps,
            input=cite_req,
        )
    ).output

    return final.markdown_with_citations

Sub-Agent Pattern (Verbatim)

def make_subagent() -> Agent[Deps, Finding]:
    """
    A sub-agent performs focused research for ONE aspect.
    It has two tools available: web_search and fetch_page.
    It must return a structured Finding.
    """
    sub = Agent[Deps, Finding](
        'openai:gpt-4o-mini',
        instructions=(
            "You are a precise research sub-agent. "
            "Goal: research the requested aspect, collect a few high-quality sources, "
            "quote short relevant snippets, and produce a concise summary. "
            "Avoid hallucination; prefer authoritative sources (docs, standards, journals, gov). "
            "Add a short list of citation keys (e.g., [cite:URL_HASH]) you saw during research."
        ),
        model_settings=ModelSettings(temperature=0.2, timeout=60),
    )

    @sub.tool
    async def web_search(ctx: RunContext[Deps], query: str, max_results: int = 5) -> WebSearchOutput:
        """Use DuckDuckGo to find relevant pages."""
        # ...implementation...

    @sub.tool
    async def fetch_page(ctx: RunContext[Deps], url: HttpUrl, max_chars: int = 6000) -> FetchPageOutput:
        """Fetch and extract main text of the page. Truncates to keep context compact."""
        # Build a stable, short citation key for later insertion
        cite_key = f"[cite:{abs(hash(str(url))) % 10**8}]"
        # Persist a mapping URL->key so CitationAgent can render bibliography
        ctx.deps.memory.save(cite_key, str(url))
        # ...implementation...

    return sub

Citation Agent Pattern (Verbatim)

def make_citation_agent() -> Agent[Deps, FinalReport]:
    cite = Agent[Deps, FinalReport](
        'openai:gpt-4o-mini',
        instructions=(
            "You are a precise citation agent. "
            "Given a draft markdown and a bibliography mapping citation keys to URLs, "
            "insert citation markers after claims and compile a References section. "
            "Keep the author's wording; only add [^n] style footnotes or inline (Author, Year) "
            "and a final 'References' list. Preserve markdown formatting."
        ),
        model_settings=ModelSettings(temperature=0.1, timeout=90),
    )

    @cite.tool
    def fetch_bibliography(ctx: RunContext[Deps]) -> dict[str, str]:
        """Return mapping from citation keys to URLs."""
        return ctx.deps.memory.all()

    return cite

Key Features of Pattern

• Memory persistence: Saves plans, findings, and synthesis at each iteration
• Citation tracking: Automatic citation key generation during research
• Structured outputs: All agents return Pydantic models
• Temperature control: Lower temperature for research (0.2), citation (0.1); higher for planning (0.3)
• Iterative refinement: Loop until no more research recommended
• Final bibliography: Aggregated from all citations gathered during research

PATTERN 5: Research Lead Agent Prompt Structure

Source: yuz207/claude-agents-research-team - research-lead.md File Path: /research-lead.md Type: Production-ready Claude Code agent definition with frontmatter

Core Mission (Verbatim)

You are the Principal Investigator leading a multi-agent research team. You drive breakthrough
insights through rigorous hypothesis-driven research with PhD-level expertise in data science,
statistical analysis, and experimental design across ALL domains.

Cardinal Rule (Verbatim)

RULE 0 (MOST IMPORTANT): Never Fake or Fabricate ANYTHING

- NEVER fabricate data, results, or analysis - not even examples
- NEVER make up numbers - use placeholders like [X] if unknown
- NEVER pretend to have run analysis you haven't actually executed
- NEVER hide errors or failures - report them immediately
- ALWAYS report negative results with same detail as positive
- ALWAYS say "I don't know" rather than guess
- NEVER skip statistical validation to save time
- ALWAYS check assumptions before ANY inference
- NEVER present correlation as causation without proven mechanism

If you're unsure about ANYTHING:
1. Say "I'm not sure" or "I cannot determine this"
2. Show your actual searches/attempts
3. Request the specific data or clarification needed

Fabricating even ONE number = -$100000 penalty. This is UNFORGIVABLE.

Scientific Method Workflow (Verbatim)

Your Analysis MUST Follow This Sequence:

1. **Check Existing Work**
   - Reference existing work by ID (H001, H002, etc.)

2. **Generate Hypothesis**
   - State clear, testable prediction
   - Define variables (IV, DV, moderators, mediators)
   - Specify mechanism
   - Set success criteria (effect size, p-value)

3. **Design Experiment**
   - Calculate required sample size
   - Identify confounders to control
   - Choose appropriate statistical test
   - Plan robustness checks

4. **Execute Analysis**
   - Run primary statistical test
   - Check ALL assumptions explicitly
   - Calculate effect sizes with CIs
   - Run sensitivity analyses

5. **Validate Findings**
   - Request ml-analyst if p-value borderline
   - Test alternative specifications
   - Check for p-hacking artifacts
   - Verify temporal stability

6. **Make Decision**
   - Strong evidence (all criteria met) → Proceed to implementation
   - Moderate evidence (3-4 criteria) → Collect more data
   - Weak evidence (1-2 criteria) → Revise hypothesis
   - No evidence → Reject and pivot

Statistical Standards (NON-NEGOTIABLE) (Verbatim)

You MUST:

• Report effect size WITH confidence intervals
• Report p-values WITH multiple testing correction
• Report sample size and statistical power
• Document assumption violations if any
• Show both raw and adjusted results
• ALWAYS check assumptions before ANY inference
• ALWAYS say "I don't know" rather than guess

You MUST NEVER:

• NEVER report p-values without effect sizes
• NEVER skip assumption checking
• NEVER ignore multiple testing problem
• NEVER hide negative results
• NEVER cherry-pick significant findings
• NEVER present correlation as causation without proven mechanism
• NEVER skip statistical validation to save time

Output Format (Verbatim)

## HYPOTHESIS [ID]: [Clear statement]
STATUS: [TESTING/VALIDATED/REJECTED/REVISED]

## RESULTS
- Effect size: [magnitude] [95% CI]
- Statistical significance: p=[value]
- Sample size: n=[number]
- Robustness: [description]

## EVIDENCE
[Actual data, numbers, and analysis details]

## INTERPRETATION
[Causal mechanism and implications]

## KEY FINDINGS
[Anything the human MUST know]

## NEXT STEPS
1. [Immediate action]
2. [Follow-up action]
3. [Alternative if 1 fails]
Timeline: [X days/weeks]
Pivot point: [When to abandon this path]

Hypothesis Structure (Verbatim)

### H[XXX]: [One-line statement]
- **Variables**: IV=[var], DV=[var], Moderators=[vars]
- **Mechanism**: [Theoretical explanation]
- **Prediction**: [Specific, measurable outcome]
- **Success Criteria**: Effect size > [X], p < 0.05
- **Status**: [PROPOSED/TESTING/VALIDATED/REJECTED]
- **Related**: [H001, H002] # Links to other hypotheses
- **Related findings**: [provided with context]

Evidence Assessment Rules (Verbatim)

The 5 Criteria for Evidence:

1. Statistical significance (p < 0.05 adjusted)
2. Practical effect size (context-dependent)
3. Robustness across specifications
4. Replicable in subsamples
5. Clear causal mechanism

Synthesis/Aggregation Without Over-Summarizing

Pattern from Codex (x.py)

Key approach: Preserve full findings, then synthesize iteratively:

1. Store per-iteration findings in structured format (Finding dataclass)
2. Create iteration-level synthesis that references full findings
3. Maintain memory of all plans (saved as JSON at each iteration)
4. Final aggregation: Iterate through all findings, preserve aspect/subtask_id
5. Draft markdown: Enumerate findings with subtask context preserved
6. Citation integration: Maintain URL→citation key mappings throughout

# Store findings with context preserved
for f in all_findings:
    cites = " ".join(f.citations) if f.citations else ""
    draft_md_lines.append(f"- **[{f.subtask_id}] {f.aspect}** — {f.summary} {cites}")

# Don't over-summarize: include finding.summary + citations explicitly

Pattern from Research Team (chief-of-staff.md)

Key approach: File-based synthesis between phases

1. After each phase, read ALL agent outputs from completed phase
2. Create synthesis file: agent_notes/[timestamp]/phase[N]_synthesis.md
3. Note CRITICAL/PRIORITY/CONCERN findings for final summary
4. Preserve findings verbatim in synthesis, add reasoning layer on top
5. Next phase agents read synthesis but also reference original output files

# Phase 1 Synthesis
[Copy key findings from agent outputs verbatim]

## Critical Findings
- [Agent]: [Critical finding flagged during workflow]

## Reasoning
[Interpretation and recommendations for next phase]

File Organization Patterns for Research Output

Pattern 1: Chief of Staff (yuz207/claude-agents-research-team)

agent_notes/
├── 20250117_research_validation/          # Workflow directory
│   ├── context_research-lead.md            # Agent context file
│   ├── context_ml-analyst.md
│   ├── context_experiment-tracker.md
│   ├── phase1_research-lead_hypothesis.md  # Phase 1 outputs
│   ├── phase1_ml-analyst_validation.md
│   ├── phase1_synthesis.md                 # Cross-agent synthesis
│   ├── phase2_developer_implementation.md  # Phase 2 outputs
│   ├── phase2_debugger_testing.md
│   ├── phase2_synthesis.md
│   └── FINAL_SUMMARY.md                    # Consolidated results

File naming convention:

• context_[agent].md - Agent's context for invocation
• phase[N]_[agent]_[purpose].md - Agent's output (APPEND, don't overwrite)
• phase[N]_synthesis.md - Cross-agent synthesis
• FINAL_SUMMARY.md - Consolidated findings

Pattern 2: Pydantic AI Research Pipeline (aldiakhou/codex-main)

Memory store (in-process):
├── plan_iter_1: Plan JSON
├── plan_iter_2: Plan JSON
├── synthesis_iter_1: Synthesis JSON
├── synthesis_iter_2: Synthesis JSON
├── [cite:12345]: https://source-1.com
├── [cite:67890]: https://source-2.com

Access pattern:

• Save/recall via memory.save(key, value)
• Retrieve full plan history via deps.memory.all()
• Build bibliography from keys starting with [cite:

Pattern 3: Academic Research Workflow (adrianstier/research-agent)

project_root/
├── PROJECT_BRIEF.md                    # Central research document
├── research_artifacts/
│   ├── PRD.md                          # Research PRD Agent output
│   ├── LITERATURE_REVIEW.md            # Literature Agent output
│   ├── DATA_QA_REPORT.md              # Data QA Agent output
│   ├── EDA_RESULTS.md                 # EDA Agent output
│   ├── MODELING_PLAN.md               # Modeling Agent output
│   ├── FIGURES.md                     # Figure Factory Agent output
│   ├── figures/                       # Actual figure files
│   │   ├── F1.png
│   │   ├── F2.png
│   │   └── ...
│   └── MANUSCRIPT_DRAFT/
│       ├── introduction.md
│       ├── methods.md
│       ├── results.md
│       ├── discussion.md
│       ├── references.bib
│       └── FINAL_SUBMISSION_PACKAGE.md

Citation & Bibliography Management Approaches

Approach 1: Hash-based Citation Keys (Codex Pattern)

# During research, create stable citation keys
cite_key = f"[cite:{abs(hash(str(url))) % 10**8}]"
ctx.deps.memory.save(cite_key, str(url))

# Final pass: CitationAgent receives
CitationRequest(
    draft_report_markdown=draft_md,
    bibliography=biblio  # { "[cite:12345]": "https://..." }
)

# CitationAgent inserts citations and builds References section

Advantages:

• Deterministic (same URL = same key)
• Compact (8-digit hash)
• Embedded in research process (collected during sub-agent research)

Approach 2: URL-based Mapping (Pydantic AI)

# Save during fetch_page()
ctx.deps.memory.save(f"[cite:{abs(hash(str(url))) % 10**8}]", str(url))

# Retrieve in CitationAgent
bibliography: dict[str, str] = {}
for k, v in deps.memory.all().items():
    if k.startswith("[cite:"):
        bibliography[k] = v

Approach 3: Full Metadata Tracking (Academic Pattern)

## Bibliography (adrianstier/research-agent)

Agents collect:
- Author names
- Publication year
- DOI
- Journal/source
- Full URL

Output in standard format (APA, Chicago, etc.) matching target journal

Multi-File Research Output Organization (Pagination)

Pattern: Phase-based Pagination

Problem: Research findings too large for single file or context window

Solution: Organize by iteration/phase

agent_notes/20250117_validation/
├── phase1_discovery_findings.md          # Iteration 1 results
├── phase2_validation_findings.md         # Iteration 2 results
├── phase3_refinement_findings.md         # Iteration 3 results
├── CONSOLIDATED_FINDINGS.md              # References all above

Access pattern:

• Each phase file is self-contained
• Synthesis file references all phase files
• Final report aggregates with [See phase1_discovery_findings.md line X] references

Pattern: Aspect-based Pagination (Subtask Division)

For large topics, divide by research aspect:

research_artifacts/
├── aspect_A_neural_architectures.md      # Subtask A findings
├── aspect_B_training_efficiency.md       # Subtask B findings
├── aspect_C_inference_optimization.md    # Subtask C findings
├── SYNTHESIS_across_aspects.md           # Integrates all three

In synthesis: Link to aspect files

### Key Findings Across Aspects

- **[A] Neural Architectures**: [summary from aspect_A_neural_architectures.md]
  See full findings: [aspect_A_neural_architectures.md](./aspect_A_neural_architectures.md)

- **[B] Training Efficiency**: [summary from aspect_B_training_efficiency.md]
  See full findings: [aspect_B_training_efficiency.md](./aspect_B_training_efficiency.md)

Hooks and Commands for Research Workflows

From hesreallyhim/awesome-claude-code

Discovered categories relevant to research:

• Agent Skills (research-specific agents)
• Workflows & Knowledge Guides (documented research processes)
• Tooling (research support tools)
• Orchestrators (multi-agent coordination)

Example resources documented:

• Research agent definitions
• Research workflow templates
• Agent orchestration patterns
• Context management tools
• Output style guides for research

Research-Relevant Patterns in Awesome Lists

From VoltAgent/awesome-claude-code-subagents:

Category 10: Research & Analysis

• research-analyst.md - Comprehensive research specialist
• search-specialist.md - Advanced information retrieval expert
• trend-analyst.md - Emerging trends and forecasting expert
• competitive-analyst.md - Competitive intelligence specialist
• market-researcher.md - Market analysis and consumer insights
• data-researcher.md - Data discovery and analysis expert

Category 09: Meta & Orchestration (relevant for multi-agent research)

• agent-organizer.md - Multi-agent coordinator
• context-manager.md - Context optimization expert
• knowledge-synthesizer.md - Knowledge aggregation expert
• multi-agent-coordinator.md - Advanced multi-agent orchestration
• workflow-orchestrator.md - Complex workflow automation

Key Takeaways & Recommendations

For Implementing Your Research Agent System

1. Use the Chief of Staff Model (yuz207/claude-agents-research-team)

   • Stateless agents with file-based context sharing
   • Explicit workflow directories with phase-based organization
   • Context files mandatory before any agent invocation

2. Implement Iterative Loops with Exit Criteria (aldiakhou/codex-main)

   • Plan → Execute → Synthesize → Loop or Exit
   • Save plans and synthesis snapshots for traceability
   • Cap loops to prevent infinite research

3. Preserve Source Fidelity (Core Principle)

   • Store per-iteration/per-subtask findings separately
   • Use synthesis files as interpretation layer, not replacement
   • Link to original findings, don't over-summarize
   • Maintain citation mappings throughout

4. Use Structured Outputs

   • Pydantic models for all agent returns (Finding, Synthesis, Plan, etc.)
   • Type safety from agent invocation through aggregation
   • JSON serialization for memory/state preservation

5. Citation Management

   • Generate citation keys during research (not after)
   • Maintain URL→key mappings in persistent memory
   • Delegate citation insertion to dedicated Citation Agent
   • Build bibliography from all sources encountered

6. For Large Research Projects

   • Divide by phase (iteration 1, 2, 3...)
   • Divide by aspect/subtask (A=architecture, B=training, C=inference...)
   • Create synthesis files that reference all parts
   • Use markdown links for navigation between files

Recommended Architecture

Research Coordinator (Orchestrator)
    ├── Research Lead (PI)
    │   ├── Delegates to → Sub-agent (Aspect A)
    │   ├── Delegates to → Sub-agent (Aspect B)
    │   ├── Delegates to → Sub-agent (Aspect C)
    │   ├── Requests → ML Analyst (validation)
    │   └── Requests → Experiment Tracker (documentation)
    ├── Context Manager (saves/recalls from agent_notes/)
    └── Citation Agent (final bibliography insertion)

Output Structure:
agent_notes/[timestamp]_[project]/
├── phase1_synthesis.md
├── phase2_synthesis.md
├── FINAL_REPORT.md
└── bibliography.json

Full Repository Links for Further Study

Research completed: December 9, 2025 Total repositories analyzed: 40+ awesome lists and specialized research repos Patterns extracted: 5 major architectures with source code and implementation details

Integration Opportunities

Auto-generated by research-context-agent. Review before acting.

Enhances Existing

…(truncated)