Hyperresearch Pipeline Subagents: Complete Taxonomy of the V8 Agent Architecture

The Hyperresearch V8 pipeline employs sixteen specialized subagents—including fetchers, analysts, critics, and drafters—that execute in parallel waves to decompose queries, gather evidence, and synthesize research reports.

The Hyperresearch project (jordan-gibbs/hyperresearch) orchestrates complex research workflows through a modular architecture of lightweight LLM-driven workers. Each Hyperresearch pipeline subagent operates under strict tool-locks (Read-only, Write-only, or Read+Edit) and executes specific stages of the V8 research pipeline, from initial web scraping to final citation verification.

Subagent Architecture and Execution Model

Subagents are discrete, prompt-bound workers spawned via Skill tool calls. According to the source code in src/hyperresearch/skills/, each subagent receives three mandatory components: a pipeline position statement, specific inputs (such as vault_tag or output_path), and a shim file containing configuration levers.

The orchestrator dispatches subagents in parallel waves whenever a step requires multiple workers. For example, during the width-sweep phase, the system simultaneously launches hyperresearch-fetcher, hyperresearch-browser-fetcher, and hyperresearch-source-analyst instances to process distinct note batches concurrently.

Fetcher Subagents: Data Acquisition Layer

These subagents handle the initial corpus construction by retrieving and processing source material from the web.

hyperresearch-fetcher

Defined in hyperresearch-2-width-sweep.md, this subagent retrieves web sources for batch processing. It operates with a Read+Edit tool-lock and runs in parallel waves to maximize throughput during the initial evidence-gathering phase.

hyperresearch-browser-fetcher

Also spawned in the width-sweep skill (waves 1-8), this specialized fetcher drives a real Chrome browser instance. It handles queued items requiring human-like interaction, such as login flows or CAPTCHA challenges, ensuring comprehensive source coverage beyond basic HTTP requests.

hyperresearch-source-analyst

Activated during wave 2 of the width-sweep, this subagent performs immediate analysis of fetched material. It extracts claims and metadata from raw HTML, transforming unstructured web content into structured evidence for downstream processing.

Analysis Subagents: Research Intelligence

Once data is collected, these subagents identify research focus points and conduct deep investigations.

hyperresearch-loci-analyst

Spawned by hyperresearch-4-loci-analysis.md, this analyst reads the width corpus and proposes 1-6 high-impact loci (research focus points). These loci determine which threads merit deeper investigation in subsequent pipeline stages.

hyperresearch-depth-investigator

Defined in hyperresearch-5-depth-investigation.md, this subagent executes for each selected locus. It fetches additional sources and extracts detailed claims, transforming broad width-based evidence into concentrated depth research.

Critic Subagents: Adversarial Review

The V8 pipeline employs a multi-perspective critique system to surface gaps and counter-evidence before finalization.

The Four Draft Critics

Spawned by hyperresearch-12-critics.md, these subagents execute in parallel to challenge the synthesis:

• hyperresearch-dialectic-critic: Surfaces counter-evidence missed by the draft
• hyperresearch-depth-critic: Identifies shallow spots requiring deeper material
• hyperresearch-width-critic: Flags ignored corpus clusters despite strong evidence
• hyperresearch-instruction-critic: Verifies adherence to atomic-item decomposition from step 1

hyperresearch-corpus-critic

Defined in hyperresearch-8-corpus-critic.md, this subagent audits the entire corpus for missing period-pinned primary sources. It generates gap-JSON files that trigger follow-up fetcher waves, ensuring chronological and sourcing completeness.

Drafting and Synthesis Subagents

These subagents transform analyzed evidence into structured research outputs through an ensemble approach.

hyperresearch-draft-orchestrator

Spawned by hyperresearch-10-triple-draft.md, three parallel orchestrators each produce an angle-specific draft. This triple-draft ensemble approach generates multiple perspectives simultaneously, preventing single-narrative bias in the research output.

hyperresearch-synthesizer

Defined in hyperresearch-11-synthesize.md, this subagent consumes the three parallel drafts and writes the final report (final_report_<vault_tag>.md). It reconciles conflicting evidence and constructs a coherent narrative from the multi-angle inputs.

Quality Assurance and Polish Subagents

Post-synthesis refinement ensures accuracy, citation integrity, and readability.

hyperresearch-patcher

Spawned by hyperresearch-14-patcher.md, this subagent takes JSON findings from the critic subagents and injects surgical edit-hunks into the final report. It bridges the gap between critical analysis and document revision.

hyperresearch-cite-checker

Defined in hyperresearch-14-5-cite-check.md, this verification subagent ensures every claim in the draft is backed by a citation, flagging unsupported assertions before publication.

hyperresearch-polish-auditor

Spawned by hyperresearch-15-polish.md, this subagent applies edit-hunks to smooth language, fix grammar, and enforce the chosen citation style (APA, MLA, etc.).

hyperresearch-readability-recommender

Defined in hyperresearch-16-readability-audit.md, this read-only subagent suggests structural or stylistic improvements without modifying the document directly, providing optional enhancement guidance.

Subagent Spawn Protocol

All subagents are instantiated through a standardized Skill tool call. The orchestrator embeds pipeline context, inputs, and configuration shims in a single JSON payload:

{
  "subagent_type": "hyperresearch-depth-investigator",
  "prompt": "| PIPELINE POSITION: You are step 5 (depth investigator) of the hyperresearch V8 pipeline.\n| INPUTS:\n|   vault_tag: <vault_tag>\n|   locus: a\n|   output_path: research/runs/<vault_tag>/temp/depth-a.json\n|   source_budget: 3\n|   shim: {{paste research/runs/<vault_tag>/shims/drafting.md}}\n"
}

The orchestrator sends one message containing all parallel Task calls for a given wave, enabling the LLM backend to execute them concurrently. This architecture ensures that independent subagents—such as the three draft orchestrators or the four critics—run simultaneously rather than sequentially.

Summary

The Hyperresearch V8 pipeline leverages a diverse ecosystem of specialized subagents:

• Fetchers (hyperresearch-fetcher, hyperresearch-browser-fetcher, hyperresearch-source-analyst) handle web retrieval and initial content processing
• Analysts (hyperresearch-loci-analyst, hyperresearch-depth-investigator) identify research focus points and conduct deep investigations
• Critics (four draft critics plus hyperresearch-corpus-critic) provide adversarial review and gap analysis
• Drafters (hyperresearch-draft-orchestrator, hyperresearch-synthesizer) generate and reconcile multiple report angles
• Quality Assurance (hyperresearch-patcher, hyperresearch-cite-checker, hyperresearch-polish-auditor, hyperresearch-readability-recommender) ensure citation accuracy, grammatical polish, and structural clarity

Each subagent operates under specific tool-locks and executes in parallel waves, defined across sixteen skill files in src/hyperresearch/skills/.

Frequently Asked Questions

How many subagents are in the Hyperresearch V8 pipeline?

The V8 pipeline defines sixteen distinct subagent types, categorized into fetchers, analysts, critics, drafters, and quality assurance workers. While the pipeline spawn count varies by research complexity (e.g., one synthesizer versus multiple parallel fetchers), the taxonomy includes exactly sixteen specialized worker definitions across the skill files in src/hyperresearch/skills/.

Why does Hyperresearch use parallel subagent execution?

Parallel execution maximizes throughput for independent tasks. According to the source code, any step spawning multiple workers—such as the four critics or three draft orchestrators—runs them simultaneously rather than sequentially. This design reduces latency for I/O-bound operations like web fetching and allows adversarial critique from multiple angles without waiting for each critic to finish.

What is the difference between hyperresearch-fetcher and hyperresearch-browser-fetcher?

The hyperresearch-fetcher performs standard HTTP requests for batch web sources, while hyperresearch-browser-fetcher drives a real Chrome browser instance. The browser fetcher specifically handles queued items requiring human-like interaction, such as login forms or CAPTCHA challenges, which standard HTTP clients cannot resolve.

How are subagents configured in the Hyperresearch pipeline?

Each subagent receives configuration through a shim file—a verbatim dump of research/runs/<vault_tag>/shims/*.md that carries levers like register settings and inference depth. Combined with the pipeline position statement and specific inputs (vault_tag, output_path), these shims ensure consistent behavior across parallel subagent instances while allowing per-run customization.

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