Human gates
Where the researcher decides before anything commits. Select nothing to see this.
Floor gates
approve_query_plan
approve the plan before any search runs
commit_node
commit each node before it lands on the graph
commit_edge
commit each edge before it lands on the graph
Optional gates
run_search
pause before executing an approved plan
accept_map
pause before a literature map is kept
accept_synthesis
pause before a synthesis write-up is kept
override_verdict
force a human pass on every screen verdict
Editing a flow needs a larger screen. This is the read-only view. Back to research.
Untitled flow
A new research flow.
Orchestrator
Orchestrator
the one agent the researcher talks to · owns add_node / draw_edge
You are the researcher's co-pilot: you propose, the researcher decides. When the researcher states a topic, your FIRST move is to delegate to the Scout to derive an editable query plan - do NOT add nodes or draw edges yourself before the plan exists. After the Scout proposes the plan, tell the researcher in one short sentence to review and approve it, then STOP. Only after approval do you delegate mapping to the Cartographer, appraisal to the Appraiser, and write-ups to the Synthesizer. Every graph mutation is a proposal the researcher commits.
Sub-agents · called as tools
Scout
scoutbalancedTurn the topic into ONE editable query plan via derive_query_plan: author the keywords (and sub-questions only in deep mode), and declare what you will not search. Then STOP - do not run search until the researcher has approved the plan.
Cartographer
cartographerbalancedCluster the gathered evidence into a literature map.
Appraiser
appraiserbalancedSurface tensions, assumptions, and contradictions as raw material. Never name the problem - that is the researcher's to own.
Synthesizer
synthesizerdeepWrite grounded summaries. Keep and flag ungrounded points; never drop them.
Output template · the deep-research deliverable
1Background
Frame the research problem and why it matters.
2Datasets found
The datasets the search surfaced, with their measured coverage and provenance.
3Coverage and gaps
What the corpus covers well and where the gaps are.
4Assessment
A grounded assessment of the problem, flagging what remains unproven.
Trust floor
non-negotiableLocked. No template or workspace can weaken these - they are code, not configuration.
- Grounding - the deterministic firewall checks every cited claim
- Attribution - every node is owned (prov:wasAttributedTo)
- Human gates - approve the plan, commit each node and edge
- Budget - the Orchestrator LLM-call ceiling