Meet the Research Crew: Agentic GenAI Workflow

Roles & connections

The crew & interactions

The researcher leads and approves decisions. Compass advises; connections show information and feedback. Initiative describes when each agent starts work.

Dark two-way line: direct researcher–specialist interaction. Teal two-way arrows: information, feedback and clarification.

Five agents · individual roles

About the crew

Explore each agent’s role and the components that shape its work.

Shared example · fictional

One research project, five complementary roles

The question: How do university teachers describe adopting GenAI in higher education teaching, and what support do they need?

The material: An initial set of 18 interviews with university teachers, a literature library and a manuscript in progress.

The next step: You have completed two more interviews, bringing the study to 20. Their transcripts are ready to join the analysis. In this fictional example, the new teachers discuss using GenAI for assessment and the support they would find useful. You want to explore how these accounts enrich the existing themes and decide what to include in the manuscript.

Follow the crew: Compass helps you plan this next stage; Analyst proposes how to integrate interviews T19 and T20; Librarian finds relevant scholarship on GenAI and assessment; Scribe helps bring the approved insights into the manuscript when you ask; Editor reviews how clearly the expanded study is presented. You decide the interpretation and approve the changes.

This teaching scenario is inspired by your research interests. All interview excerpts, findings and source records below are invented; they do not describe your actual data. This page demonstrates agent designs; it does not run agents or connect to tools.

Shared setup

The crew is built from three connected layers: an interface for the researcher, a workflow that coordinates the roles, and a model that produces each response.

Roles & model: An agent role is not a separate AI brain. Compass, Librarian, Analyst, Scribe and Editor can all use the same model, while receiving different instructions, evidence, tools, permissions and conversation history. Separate models are optional.
Model runtime: LM Studio or Ollama loads the chosen model and makes it available through a local API. When another component sends a prompt, the runtime passes it to the model and returns the generated response. LM Studio also includes its own desktop chat, which is useful for testing a model or prompt—but it does not coordinate the five roles.
Workflow controller: LangGraph implements the crew’s working process. It receives the researcher’s request, selects the relevant role, adds that role’s instructions and approved context, calls the model through LM Studio or Ollama, records the result, and decides what may happen next. It can call tools, hand work to another role, preserve project state and pause until the researcher approves an action.
User interface: Chainlit is the screen the researcher uses. It sends messages and approvals to LangGraph, then displays each role’s contribution as a visible conversation. Chainlit does not run the model or decide the workflow. LibreChat is an alternative general chat interface, but offers less control over a purpose-built crew experience.
Where data runs: Local: identifiable interviews, unpublished drafts and private project summaries. Cloud: approved public sources and non-sensitive tasks, where policy permits. Hybrid: online discovery followed by local screening or analysis. Keep private retrieval, workflow state and logs inside the approved boundary.

Local model sizing & implementation notes

Qwen3 14B Q4_K_M (~9.3 GB download) is a candidate to test on 24–32 GB unified-memory machines; Qwen3 32B Q4_K_M (~20 GB) on 48–64 GB machines. These are starting points, not guaranteed requirements; context length, concurrent runs, GPU memory and CPU offload also affect feasibility.

Begin with one active role at a time and a modest context window. Evaluate task quality, evidence fidelity, tool-call reliability, latency, memory and swap. Pin the model, runtime, prompt and source versions. Record proposals, approvals, tool events and applied changes separately before adding scheduled activity.

Human-led processes

Outside the crew

These processes stay human-led: you initiate each task and use AI as a manual tool or assistant. The examples below are workflows, not additional agents. Their position in the next grid reflects how these tools are used here, rather than everything a product might be capable of.

Agents & human-led processes

Autonomy & initiative

Execution autonomy means how independently the configured workflow proceeds once started: with human guidance at each stage, or independently within agreed limits.
Initiative autonomy means how much starts without a fresh human request.

Filled circles: crew agents. Outlined circles: human-led processes outside the crew. Positions are illustrative workflow choices, not measured model scores. Click a marker to explore it.

Reading the grid: Positions show intended autonomy; the researcher retains oversight and decisions.