Research Framework
The Research Agent's structured multi-phase pipeline — core components, mission flow, search adapters, and current limitations.
Purpose
The Research Framework is the structured pipeline that governs how the Research Agent executes missions. It defines how a task is received, how research phases are executed in dependency order, how intermediate data is stored, and how a final report is produced.
This page documents the Research Agent's concrete pipeline on top of OpenClaw. The OpenClaw framework provides the general mission runtime and artifact model; this research framework defines the domain-specific phases used for structured information gathering and report generation.
Status: Implemented (pipeline, phase DAG, artifact store, and final report generation operational)
Core components
| Component | Responsibility |
|---|---|
| Task Layer | Receives task definitions from the file-based queue |
| Mission Runner | Creates the run directory, manages lifecycle, orchestrates phases |
| Phase DAG | Defines execution order and dependencies between phases |
| Search Adapters | Interface to external search and media sources |
| Artifact Store | Stores typed intermediate outputs per mission run |
| State Store | Tracks mission progress, phase status, and errors |
| Final Report | Aggregates all artifacts into final_report.md |
Mission flow
A research mission proceeds as follows:
- A task definition is written to the task queue directory as a JSON file.
- The Mission Runner detects the task and creates an isolated run directory.
- Phases execute in the order defined by the Phase DAG, respecting dependencies.
- Each phase reads from prior artifacts and writes its output to the run directory.
- The final phase (F6) aggregates all available artifacts into
final_report.md. - The state file is updated to reflect completion or any errors.
Phase DAG overview
F2, F4, and Y1 are independent of each other and can be treated as a parallel group after candidates are produced. See Research Mission DAG for full details.
Artifacts produced
Each mission run produces the following files:
| File | Contents |
|---|---|
task.json | Original task definition |
candidates.json | Discovered candidate items |
reviews.jsonl | Review and analysis entries per candidate |
price_analysis.json | Pricing and availability data |
youtube_transcripts.json | Video transcript content |
synthesis.json | Deep synthesis output (optional) |
final_report.md | Aggregated Markdown report |
state.json | Mission progress and error tracking |
See Research Artifacts for the full lifecycle description.
Search adapters
The current implementation includes adapters for:
- Web search (external search API)
- Media platform queries (streaming catalog integration)
- YouTube transcript extraction
Additional adapters for other research domains are planned.
A current limitation
The framework currently handles structured research well within its designed domains (discovery, reviews, pricing, transcripts). Support for broader research domains — academic papers, technical documentation, structured databases — is planned but not yet implemented.
Related pages
→ OpenClaw Framework Overview — the runtime model this pipeline runs on top of
→ Research Mission DAG — detailed phase dependency graph
→ Research Artifacts — artifact files and lifecycle
→ Research Agent — agent overview