Deep Researcher Agent

The Deep Researcher coordinates a structured research pipeline that separates source routing, planning, evidence collection, and final synthesis. It uses task subagents under an orchestrator plus reusable researcher workers created with the deepagents and LangChain libraries.

Location: src/aiq_agent/agents/deep_researcher/agent.py

For optional DeepAgents sandbox execution and operational notes, refer to Deep Research Sandbox.

Purpose

The deep path handles queries that require comprehensive investigation: multi-step research, comparative analyses, and topics that benefit from structured planning and evidence gathered from multiple sources. It produces the output shape requested by the user, including long-form reports, brief answers, tables, comparisons, predictions, and data extractions.

Internal Flow

graph TD A[Receive DeepResearchAgentState] --> B[Filter registry-mapped tools by data_sources
retain unmapped configured tools] B --> C[Build orchestrator and shared runtime] C --> D{Source router enabled?} D -->|yes| E[task: source-router-agent] E --> F[Write advisory SourceRoutingPlan
to /shared/source_routing.json] D -->|no| G[task: planner-agent] F --> G G --> H[Return structured ResearchPlan
persisted to /shared/plan.json] H --> I[Orchestrator calls run_research_batch
with planned ResearchQuery objects] I --> J[Concurrent reusable researcher workers
one worker per ResearchQuery] J --> K[Return structured ResearchNotes
and persist /shared/research_note_*.json] K --> L[Normative task: writer-agent] L --> M[Writer commits /shared/output.md
through the shared-state backend] M --> N{Current bytes match
this run's writer digest?} N -->|yes| O[Runtime loads writer output] N -->|no| V[Error: writer_output_not_committed] O --> T{Citation verification enabled?} T -->|yes| P[Verify citations against captured sources] T -->|no| Q[Skip citation verification] P --> R[Sanitize and return final Markdown] Q --> R style A fill:#e1f5fe style E fill:#e8eaf6 style G fill:#fff8e1 style J fill:#fce4ec style L fill:#e8eaf6 style R fill:#e8f5e9

Coordination and Data Handoffs

sequenceDiagram participant O as Orchestrator participant S as source-router-agent participant P as planner-agent participant B as run_research_batch participant R as Researcher workers participant T as Allowed source tools participant W as writer-agent participant X as Runtime opt enable_source_router O->>S: task(user request and context) S-->>O: /shared/source_routing.json end O->>P: task(user request and advisory route) P->>T: Ground the plan with allowed tools T-->>P: Discovery results P-->>O: ResearchPlan in /shared/plan.json O->>B: ResearchQuery objects from the plan par One invocation per independent query B->>R: Invoke reusable researcher runnable R->>T: Use preferred and fallback tools T-->>R: Source results R-->>B: Structured ResearchNotes end B-->>O: Notes plus /shared/research_note_*.json O->>W: task(plan, notes, captured sources) W-->>O: Commit /shared/output.md O-->>X: Writer completion marker X->>X: Verify current bytes against the run-local writer digest X->>X: Load, optionally verify citations, and sanitize

The orchestrator serializes dependent stages and tracks progress. It does not call source tools directly. Final synthesis is delegated exclusively to the writer; source access is delegated to the planner and researcher workers. A non-empty file is not sufficient proof of completion: the runtime accepts the report only when its exact UTF-8 bytes match the digest recorded after a successful writer mutation in the current run. Missing, stale, or modified output fails closed with writer_output_not_committed.

The commit proof is intentionally run-local and is not restored from a checkpoint after a process restart. The guarantee is one overwrite-capable shared-state backend update followed by byte-exact digest verification; it does not claim cross-provider filesystem atomicity. A resumed run without its proof therefore fails closed rather than trusting pre-existing output.

Runtime Roles

Participant	Invocation	Responsibility and output
Orchestrator	Root create_deep_agent graph	Coordinates stage order, reads the persisted plan, dispatches research batches, and normally delegates final synthesis. It has run_research_batch and helper tools, but no direct source tools.
source-router-agent	Optional DeepAgents task() subagent	Looks up the configured source catalog, chooses one advisory domain route, and writes a SourceRoutingPlan to /shared/source_routing.json. It does not research.
planner-agent	DeepAgents task() subagent	Grounds the requested answer strategy with available source tools and returns a structured ResearchPlan. The runtime persists it to /shared/plan.json.
Researcher workers	Reusable LangChain runnable invoked by run_research_batch	Each worker executes one self-contained ResearchQuery and returns structured ResearchNotes. Independent workers run concurrently up to the configured limit.
writer-agent	DeepAgents task() subagent	Reads the plan, research-note files, and captured sources; performs final-answer synthesis on the normative path; and writes /shared/output.md. It has no source-search tools and performs no new research.

The source router, planner, and writer are task subagents registered with the DeepAgents root graph. Researcher workers are different: they are invocations of one reusable, structured-output runnable behind the orchestrator-only run_research_batch tool. They do not appear as task() subagents and do not manage top-level workflow todos.

Shared State, Skills, and Execution Boundary

The shared-state and skills boxes in the architecture diagram represent runtime dependencies, not additional agents:

Boundary	Current implementation
Shared research state	The host-side StateBackend mounted at /shared/ stores the source-routing plan, research plan, structured notes, and /shared/output.md. DeepAgents graph state separately carries progress todos and file metadata.
Skill definitions	Built-in skill collections are mounted from the host at /skills/. Filesystem permissions expose only the collections assigned to a role and deny writes to the skill tree.
Sandbox workdir	When a sandbox is configured, the default filesystem route and execute tool use one provider sandbox per deep-research job. Agents within that job share the provider runtime; separate jobs receive separate sandboxes.
Inference and source tools	LLM calls, source-tool calls, credentials, orchestration state, and /shared/ remain in the AI-Q process. Only generated code and job-workspace files cross the sandbox boundary.

The shipped config_domain_routing_and_skills.yml profile assigns the research collection to researcher workers and the synthesis collection to the writer. The research collection currently includes chart generation, table analysis, forecast analysis, and lightweight calculations. The synthesis collection includes long-form and prediction report writers. A skill provides instructions; only skills that invoke execute require the optional sandbox. Modal and OpenShell implement the same provider-neutral job-scoped contract.

Data Source Boundary

DeepResearchAgentState.data_sources is a hard per-request boundary for tools mapped in data_source_registry. The registration layer filters those mapped tools before constructing the active deep-research agent. Configured tools that are not mapped to a registry source remain active:

• None makes all configured tools available.

• [] removes mapped data-source tools while retaining unmapped configured tools, including utilities.

• A populated list admits only tools mapped to those source IDs, plus unmapped configured tools, including utilities.

The optional source router receives a catalog containing only mapped sources within this boundary. Unmapped configured or utility tools do not appear in that catalog even though they remain active. Router recommendations are advisory and cannot restore a filtered-out mapped source.

The planner records exact available tool names in each ResearchQuery.preferred_tools and fallback_tools as structured guidance. Those fields do not narrow the callable tool set at runtime. Every researcher worker is bound to the full request-filtered tool set and is prompted to try the preferred and fallback tools in the recorded order.

Middleware and Tool Boundaries

The roles use middleware appropriate to their contracts:

Role	Relevant behavior
Orchestrator	DeepAgents task, todo, and filesystem support; source-routing and final-report ownership guards; tool-name validation restricted to helpers and run_research_batch; tool and model retry handling
Source router	Minimal filesystem and retry middleware; catalog lookup and write_file only; final-report mutation denied
Planner	Source capture, retries, filesystem access, todo suppression, structured ResearchPlan validation, automatic plan persistence, and final-report mutation denial
Researcher worker	Filesystem context, optional skills, summarization, source capture, retries, structured ResearchNotes validation, and final-report mutation denial
Writer	Filesystem context, optional synthesis skills, source-registry access, retries, todo suppression, overwrite-safe /shared/output.md commit, and run-local digest verification; no source-search tools

The root graph is constructed with create_deep_agent. The reusable researcher runnable is constructed separately with create_agent, which is what lets run_research_batch invoke independent queries concurrently.

State Model

DeepResearchAgentState

Field	Type	Default	Description
messages	Annotated[list[AnyMessage], add_messages]	required	Input query and conversation messages managed by the LangGraph message reducer
data_sources	list[str] or None	None	Hard per-request filter for registry-mapped source tools; unmapped configured tools remain active
user_info	dict or None	None	Authenticated user context available to prompts
tools_info	list[dict] or None	None	Available-tool metadata
todos	list[dict]	[]	Top-level progress list managed by the orchestrator
files	dict	{}	Merge-reduced virtual filesystem containing plans, notes, output, and optional parent-report context
subagents	list[dict]	[]	Status of configured DeepAgents task subagents; researcher worker invocations are not stored here
rubric	str or None	None	Optional DeepAgents rubric state
clarifier_result	str or None	None	Clarification log containing missing context or requested output-shape preferences gathered before research
available_documents	list[AvailableDocument] or None	None	User-uploaded documents and summaries available as research context

When present, clarifier_result is injected as context. The planner independently creates the ResearchPlan inside the deep-research workflow.

Configuration

The architecture is configured through DeepResearchAgentConfig (NeMo Agent Toolkit type name: deep_research_agent). The workflow-shaping parameters are summarized here; refer to the Configuration Reference for configuration details.

Parameter	Type	Default	Description
orchestrator_llm	LLMRef	required	LLM for workflow coordination
source_router_llm	LLMRef or None	None	LLM for optional advisory routing; falls back to orchestrator_llm
planner_llm	LLMRef or None	None	LLM for structured planning; falls back to orchestrator_llm
researcher_llm	LLMRef or None	None	LLM used by every researcher worker; falls back to orchestrator_llm
writer_llm	LLMRef or None	None	LLM for final synthesis; falls back to orchestrator_llm
tools	list[FunctionRef | FunctionGroupRef]	[]	Explicit source tools; an empty list inherits tools from the data-source registry
exclude_tools	list[str]	[]	Tool names removed from inherited tools
domain_catalog_path	str or None	None	Optional YAML or JSON domain catalog used by the source router
enable_source_router	bool	true	Run the advisory source-router stage before planning
max_research_concurrency	int	6	Maximum ResearchQuery items accepted and run concurrently per batch call
skills	FunctionRef, inline deep_research_skills, or None	None	Optional built-in skill assignments by agent name
sandbox	FunctionRef, inline deep_research_sandbox, or None	None	Optional sandbox profile for DeepAgents execute support
enable_citation_verification	bool	true	Verify generated citations against captured sources after final report extraction
resource_limits	DeepResearchResourceLimits	hard ceilings	Per-job request, graph-time, plan, report, shared-state, note, todo, query, and source-call budgets; configurable downward only
verbose	bool	true	Enable detailed logging

Example YAML:

functions:
  deep_research_agent:
    _type: deep_research_agent
    orchestrator_llm: nemotron_ultra_llm
    source_router_llm: nemotron_ultra_llm
    planner_llm: nemotron_ultra_llm
    researcher_llm: nemotron_ultra_llm
    writer_llm: nemotron_ultra_writer_llm
    enable_source_router: true
    enable_citation_verification: true
    max_research_concurrency: 6
    resource_limits:
      max_research_queries: 20
      max_source_tool_calls: 100
    verbose: true
    tools:
      - web_search_tool

The researcher filesystem view treats /shared/** as read-only. Researchers return one schema-validated ResearchNotes object per accepted ResearchQuery; the parent run_research_batch tool alone applies note count/byte quotas and persists the note. Planner persistence and top-level todo replacement are similarly centralized and validated before shared-state mutation. See the Configuration Reference for all enforced ceilings.

Note

Hosted Endpoint Availability: The default deep researcher uses Nemotron 3 Ultra (nvidia/nemotron-3-ultra-550b-a55b) for every role, including writer_llm. Shared hosted endpoints can have limited availability during high demand (HTTP 429/503 responses). For production deployments requiring consistent throughput, refer to the self-hosting guidance.

Prompt Templates

Located in src/aiq_agent/agents/deep_researcher/prompts/:

Template	Purpose
orchestrator.j2	Coordinates the ordered router, planner, batch-research, and writer handoffs
source_router.j2	Selects an advisory route from the allowed source catalog and writes SourceRoutingPlan
planner.j2	Builds the structured ResearchPlan and makes the final ResearchQuery tool choices
researcher.j2	Executes one ResearchQuery and returns structured ResearchNotes
writer.j2	Reads persisted artifacts and captured sources, then performs final synthesis into /shared/output.md

Workflow Phases

Phase 1: Advisory Source Routing

When enable_source_router is true, source-router-agent selects one domain route and source ordering from the already-allowed source catalog. It writes /shared/source_routing.json. Planning continues without this stage when it is disabled, and the route does not override the user’s source selection.

Phase 2: Research Planning

The planner uses available source tools to ground a structured ResearchPlan containing:

• Task analysis and the intended answer shape

• Required answer components and constraints

• Self-contained ResearchQuery objects

• Preferred and fallback tool guidance for each query

The planner reads source-routing guidance when available and records the final query-level tool preference order. This is prompt guidance rather than runtime tool enforcement. The runtime persists the validated plan to /shared/plan.json.

Phase 3: Concurrent Evidence Collection

The orchestrator passes the plan’s independent ResearchQuery objects to run_research_batch. The tool invokes one reusable researcher worker per query concurrently, bounded by max_research_concurrency. Each worker:

1. Reads the relevant plan context

2. Is prompted to follow the query’s preferred and fallback tool order while remaining bound to the full request-filtered tool set

3. Returns validated ResearchNotes with findings, sources, gaps, and an evidence judgment

The batch tool returns the notes to the orchestrator and persists them under /shared/research_note_*.json. If only part of a batch fails, successful notes remain registered and persisted; only failed or missing queries are eligible for another call.

Phase 4: Writer-First Final Synthesis

The orchestrator delegates once the plan and research notes are available. The writer reads /shared/plan.json, all research-note files, and the captured source registry. It may also read parent-report context for a report edit. The writer performs no new research, writes the complete final answer to /shared/output.md, and returns a short completion marker. The runtime loads the Markdown from that file.

This is the only synthesis contract. The runtime accepts only non-empty writer output whose exact UTF-8 bytes match the digest recorded after a successful writer mutation in the current run. After one bounded corrective turn, missing, stale, or mismatched output fails closed with writer_output_not_committed; inline orchestrator messages are not salvaged as final reports.

/shared/output.md is the sole writer-facing path. When CompositeBackend routes /shared/ through StateBackend, raw graph state may represent that file under the internal route-stripped key /output.md. Ownership and digest checks recognize that internal alias, but agents must not target it directly.

Phase 5: Citation Verification (Post-Processing)

Citation verification is enabled by default and configurable with enable_citation_verification. When enabled, a deterministic post-processing pipeline checks citations against sources captured from configured tools. Report sanitization runs after final report extraction regardless of this setting.

Location: src/aiq_agent/common/citation_verification.py

Source Registry

During planning and research, SourceRegistryMiddleware records URLs and citation keys returned by allowed source tools in a per-session SourceRegistry. Research-note source locators identify the compact set carried forward for writer-facing citation selection.

Citation Verification

The verify_citations() function validates citations in the report against the source registry using five URL matching strategies:

1. Exact match – raw or normalized URL

2. Truncation match – report URL is a prefix of exactly one registry URL

3. Prefix match – normalized report URL is a prefix of a registry URL

4. Child-path match – report URL path is a subpath of a registry URL

5. Query-subset match – same host and path, with a subset of query parameters

Unmatched citations are removed and recorded with an audit reason. Knowledge layer citations, such as report.pdf, p.15, are matched against citation keys with lenient page-number comparison.

Report Sanitization

The sanitize_report() function removes potentially unsafe or unreliable URLs from the report body:

• Shortened URLs

• Truncated or garbled URLs

• IP-address URLs

• Non-HTTP schemes such as javascript:, data:, vbscript:, and file:

After removals, citations are renumbered to close gaps in the reference list.

Verification Result

The verification result includes:

Field	Description
verified_report	Report text after citation verification
removed_citations	Removed citations with reasons
valid_citations	Retained citations with reference numbers

Evaluation

The Deep Researcher is evaluated using the Deep Research Bench (DRB), which measures research reports using RACE and FACT metrics. Refer to Deep Research Bench for full documentation.