Complete
ResearchGraph
An autonomous research OS: thirteen agents over a dependency graph.
- Role
- Sole author
- Stack
- FastAPI · LangGraph
- Code
- Repository →
13
specialised agents in the graph
Topological
execution order, not conversational
End-to-end
survey → gap → draft → critique
How it works
- 01 Survey 4 agents read the field
- 02 Extraction 3 agents pull claims
- 03 Gap analysis 2 agents name the hole
- 04 Drafting 2 agents write
- 05 Critique 2 agents object, capped
Thirteen agents, five stages, one topological order. A node waits for its prerequisites.
Termination conditions
The first version deadlocked. These four constraints are what made it finish.
The problem
Literature review is mostly dependency management. You cannot analyse a gap before surveying the field, and you cannot draft a proposal before the gap is named. Chat-shaped agents ignore that ordering and produce confident summaries of work they read in the wrong sequence.
The approach
Thirteen specialised agents — survey, extraction, gap analysis, drafting, critique — arranged as a topological dependency graph. A node runs only once its prerequisites have produced output, so the pipeline reads before it reasons and reasons before it writes.
What broke
The critique agent kept rejecting drafts on grounds the drafting agent had no way to act on, and the graph deadlocked on a cycle of polite disagreement. Making critique emit structured, addressable objections — and capping revision rounds — turned an argument into a loop that terminates.
What worked
End-to-end runs that go from a topic to a critiqued proposal draft with the reading trail attached. It is the piece of infrastructure my own literature reviews now start from.